Henry Ajder's Recent LinkedIn Posts

Not a bad way to come back from a LinkedIn break! Last night, NVIDIA invited me to join their celebration of the UK AI ecosystem, where CEO Jensen Huang announced £2bn of funding for UK AI startups alongside Prime Minister Keir Starmer. This followed NVIDIA's commitment alongside their partners Nscale and CoreWeave to invest £11bn in UK AI infrastructure, including OpenAI's Stargate UK. With this investment, Huang's overview of the UK's AI positioning is compelling: world leading universities and research, a mature startup ecosystem and well distributed AI infrastructure. At a time where public sentiment about the UK's future is rather mixed, it was great to hear a more bullish perspective. This isn't to say it's 'smooth sailing' from now on, but it's another critical piece falling into place to give us the best shot at ensuring AI's benefits are felt by everyone throughout the country. Thanks to Hayley Grossman, Melissa Mines, Tracey Nugent and David Hogan for being such gracious hosts and to Maria Luciana Axente, Azeem Azhar, Tom Westgarth, and Tee Ganbold 🌏 for great company throughout the night.

Far too many businesses still rely on processes that are woefully outdated now that creating realistic AI generated media is trivially easy. I'm seeing a growing number of claims of low level AI powered fraud like this being shared across forums and socials, compared to the handful of low quality attempts I'd see just a couple of years ago. Yet these lower stakes examples really are the tip of the iceberg. Whether it's recording expenses, submitting evidence for insurance claims, or providing evidence in a courtroom, the assumption of baseline media authenticity is no longer viable (and arguably hasn't been for some time!) This is where content provenance technologies and 'digital nutrition labels' (see Content Authenticity Initiative/Coalition for Content Provenance and Authenticity (C2PA)) have real commercial value. By providing a secure standard for authenticating how a piece of media has been captured/created, provenance tools like Content Credentials and Truepic help businesses avoid guesswork about what might be AI generated and instead focus on trusted signal of what's authentic to inform decisions. In other words, when you can efficiently authenticate media, processes are streamlined, vulnerabilities are reduced, and businesses can get 'stuff' done quicker.

A significant and welcome move by TikTok: Users will soon have the power to reduce the amount of AI generated content that appears in their feeds. The key word here is power. AI content isn't inherently bad, but the reason it and "slop" have hit a nerve with audiences is how it has rapidly flooded our digital spaces and is hidden in plain sight amongst real 'human' generated content. Letting audiences choose the amount of AI content they see is a well balanced response that doesn't make a value judgement about AI content, but respects users' personal preferences. I don't just see this move by TikTok as a nod to social responsibility , but also as a smart business decision. Jack Dorsey's recent funding of diVine, a Vine based platform where AI generated content is banned, and Deezer's recent ad campaign fighting back against AI generated music are just a few signs of the growing momentum behind AI free spaces. As i've been saying for years now, empowering users by giving them the choice about the kind of content they want to watch and transparency about where AI content is present on the platform is the future of authentic engagement. I do sometimes enjoy a smattering of absurdist AI generated slop (it's a good way to keep track of new workflows/model capabilities!) but being able to turn off the slop spigot is a step I hope other platforms follow. https://lnkd.in/exW9yv2S

Earlier this week, I shared an urgent AI PSA with BBC News: “don't confuse AI-enhancement, which is effectively a prediction, with new knowledge,” In the last year or so, I've observed a notable increase in amateur sleuthing about fast evolving crises on platforms like X or Reddit. Some of these sleuths are using legitimate OSINT techniques, but a worrying number are also using AI enhancement, upscaling, and other generative tools to supposedly 'get to the bottom' of what really happened. But as I told Thomas C., this is a fundamental mistake: “AI-enhancement tools don’t have some privileged knowledge of the reality that lies beneath a low-quality image. Instead, they approximate what an enhanced version of an image could look like,” warns generative-AI expert Henry Ajder. In other words, these tools don't reveal reality, they infer a possible one from incomplete information. Overestimating AI's abilities is common, and sometimes understandable given the allure of a confidence sounding system that can conjure plausible sounding/looking outputs in a flash. However, like hallucinations or false positives/negatives in deepfake detection, knowledge of how these systems work and a sober awareness of what they can/cannot do is essential. It would be amazing if AI enhancement tools really could reveal hidden realities, but this capability remains firmly in the realm of fantasy...

Publication news! Proud to have co-authored this major new Demos report out today: Epistemic Security for Crisis Resilience. When our information ecosystems and ability to know what's happening in the world fail, it can trigger, accelerate, or deepen national crises. In this sense, epistemic security is national security. Back in 2018, I joined the first set of epistemic security summits hosted by The Alan Turing Institute, where deepfakes and concerns of AI distortions of the information ecosystem were just starting to emerge. Fast forward to today, and the landscape looks very different, both in anticipated and unexpected ways. In our report, led by the brilliant Elizabeth Seger at Demos and in partnership with the Centre for Emerging Technology and Security (CETaS), we forecast and predict how these crisis scenarios could materialise and take hold in this landscape- from bank runs and compromised elections to violent rioting and legal system collapse. In response, we propose seven strategic policy interventions for governments to strengthen national security and democratic institutions/processes. Huge props to my co-authors Elizabeth Seger, Sam Stockwell, Tyreese Calnan, Jamie H., and Hannah Perry who did much of the heavy lifting putting this report together. Please do have a read and share widely! https://lnkd.in/ezPc3yt8

Great to feature in the Financial Times's new documentary on deepfakes with Melissa Heikkilä, who visited me in Cambridge for an in depth interview. We had a wide ranging conversation, but one key message I shared is that while the deepfake landscape has radically changed since 2017, some key dynamics have stayed the same. Yes, progress in realism, efficiency, and accessibility make today's deepfake landscape almost unrecognisable compared to eight years ago. Yet the dynamics of how harms are caused (deception, doubt, and degradation) and the commonly understood blueprint for how we should fight back remain largely unchanged. Particularly when it comes to detecting/spotting deepfakes, an adversarial/cat and mouse dynamic often unfolds as so: - New models and tools are released into the wild - Flaws in these models' outputs are identified and communicated/accounted for by detection companies, forensic experts, and society generally. - Knowledge of these flaws is made redundant and even harmful as they are trained out of new models/model iterations, often without us immediately noticing. The challenge is that as technical advances with deepfakes have accelerated, competing in this dynamic from the detection side has become increasingly fraught. As I said to Melissa, the everyday person cannot be expected to become (and shouldn't try to be!) a 'digital Sherlock', but our digital infrastructure is also not currently designed to do the heavy lifting of confident content authentication. If we're to avoid slipping even further behind, this digital infrastructure and how trust is securely mediated need fundamentally rethinking by businesses and governments. It certainly keeping me busy! Big thanks to Gillian Tett for kindly hosting us at King's College, Cambridge and to Thomas Hannen for putting together a great film (see comments for the full documentary).

A wild measure of deepfake believability: Police are asking kids to stop sending their parents AI pranks of homeless intruders that are causing panicked 911 calls. I've seen several other formats of this viral prank, from AI images of 'sexy' plumbers and cleaners being sent to partners and spouses, to celebrities appearing in people's bedrooms. Social media has been flooded with screenshots of chats of family and friends being fooled by these AI generated visitors, to the extent that police in Salem, Massachusetts are asking kids to stop because calls from scared parents are diverting resources for policing actual crime. Besides not being in the best taste, this examples says a lot about how people react to realistic AI content/deepfakes in perceived crisis scenarios. When a parent receives an realistic image from their child of a homeless man in their bed, they have agency to respond to what they believe is real: i.e. they can, and clearly in a notable number of cases, do call the police in a panic. Yet many emotionally charged deepfakes spread about issues people have less direct agency to respond to, such as foreign wars, domestic riots, or political corruption. There isn't clear and direct action we can take to 'resolve' these situations, even if we deeply believe them to be real and feel similar reflexive anger, fear, panic etc. Reacting to an intruder in your child's home isn't the exact same scenario, but it shows how emotionally charged AI content doesn't just have the ability to shape our beliefs, but can drive behaviour when we believe we can meaningfully respond. https://lnkd.in/ezafQDhE

As conflict rages in the middle-east, deepfakes have made the 'fog of war' thicker than ever. I spoke to BBC News about why this time is different compared to the other wars i've monitored for deepfakes since 2018. The use of deepfakes to spread disinformation and doubt in war isn't new. Conflicts including the ongoing Russian invasion of Ukraine, Israel Gaza war, and brief 2025 India Pakistan crisis all saw deceptive AI generated images, videos, and audio being circulated. This was in addition to the 'evergreen' cheapfakes such as video game footage, out of context photos, and crude edits that often circulate in acute low information crisis scenarios. What's different with the current conflict, as I shared with Thomas C., is the volume and realism of AI generated content being created as powerful AI tools have become commodified. "The number of different tools that are now available to create a wide range of highly realistic AI manipulations is unprecedented," says Henry Ajder, a generative AI expert. "We have never seen these tools so available, so easy and so cheap to use," he says. The result is the volume has been overwhelming at the same time that discerning authentic from deepfake/synthetic footage has become more laborious and riddled with doubt. This isn't just for everyday people trying to navigate a choked X/Twitter feed, but also for governments and other institutions that need to make informed decisions about how to respond to fast evolving and high risk situations. Last year, I posted about how a fake AI generated images of a damaged train bridge caused UK train operators to immediately cancel services. With lives on the line, a decision was made to act with caution based on the consequences if it were real and the lack of confidence in making that determination. Unsurprisingly, military and intelligence have far better digital forensic capabilities than a train operator when assessing content authenticity, but as we saw with the tragic bombing of an Iranian girl's school, intelligence flaws exist and in the deepfake context detection tools are far from invulnerable. Floods of highly realistic deepfakes in fraught scenarios is the new norm. Now more than ever, we need better 'fog lamps' to cut through these increasingly murky times (see Coalition for Content Provenance and Authenticity (C2PA), Content Authenticity Initiative, GetReal Security, and Google SynthID) Read the full piece below, with excellent insights from WITNESS' Mahsa Alimardani, Timothy Graham, and Victoire Rio. https://lnkd.in/d6cSPpA8

A grim but important discovery: New research reveals the AI nudification app economy is making $36.4m a year, and even that's a conservative calculation. This is just one of several standout stats from Santiago L.' and Alexios Mantzarlis' investigation into how the AI nudifier ecosystem profits from the abuse of millions. It's released as Unicef report that one in six Spanish children have been targeted by AI nudification apps and schools around the world continue to sound the alarm about their use on students and teachers alike. Since I published my investigation into the first Telegram nudification bot in 2019/2020, the landscape has evolved significantly. This is one of the best pieces of work i've seen illustrating where these changes have taken place in terms of scale and sophistication. It also highlights another critical conversation that needs more attention: The role of big companies who provide the core infrastructure that allow nudification apps to function. Whether its app stores, hosting services, or payment providers, those who facilitate what I call "The Perverse Customer Journey" have often avoided the scrutiny they deserve. Addressing this global scourge is far from simple, but this research is essential to making the scale of the problem tangible and helping to hold those enabling it to account. For that, Santi and Alexios deserve immense praise. Read the full investigation in the excellent Indicator newsletter below- an essential sign-up for keeping track of AI deception and weaponisation: WIRED's Matt Burgess' excellent write-up, where I shared my thoughts, also linked below: https://lnkd.in/e9iVWVcm https://lnkd.in/euP6bxrA

A real pleasure to chair the closing of the UK Government's inaugural Deepfake Detection Challenge! Concern about the impact of deepfakes is the highest i've seen since inception in 2017, but global leadership in responding has long been lacking. This challenge, along with its accompanying SoTA dataset for benchmarking detection models, are a significant signal from the UK government that they want to take up that mantle. This wasn't a talking shop or an abstract technical exercise, but a carefully designed set of scenarios putting some of the world's leading detection tools to the test. Evaluating on metrics including robustness, explainability, and attribution, as well as providing evolving datasets for ongoing benchmarking, are essential to ensuring deepfake detection tools don't become static black boxes frozen in time. Without these evaluations and grounded education on detection, deployment in critical contexts such as law enforcement, banking, or courtrooms are a recipe for disaster. But as AI Security Institute's Andrew S, WITNESS' Mahsa Alimardani, University of Southampton's Jennifer Williams, PhD, and Coefficient's John Sandall discussed on the panel below, the future holds no room for complacency. Deepfakes and synthetic media are still evolving at breakneck speed and detection is often playing catch-up- which is no surprise given the amounts being spent researching novel generation v detection. Likewise, it's important not to see any solution approach as a silver bullet, but to consider how detection can work alongside other provenance approaches such as Content credentials (See Content Authenticity Initiative and Coalition for Content Provenance and Authenticity (C2PA) and vice versa. There's no sugarcoating the scale of the challenge, but its initiatives like these that make me hopeful we can rise to it and that government can be a rally force. Congratulations to UK Home Office's Andrew Tyeloo (without whose vision and leadership this challenge would not exist) and his team, Accelerated Capability Environment (ACE)'s Zac Ghaffar, Eloise Heyraud (nee Charig), and Iain Wallace, as well as all the teams who participated. Here's to next year!

“It's a societal scourge, and it’s one of the worst, darkest parts of this AI revolution and synthetic media revolution that we're seeing,” Ajder says. Good to speak with WIRED's Matt Burgess for this important piece on the dark evolution of AI nudification technology, where I reflect on my findings from mapping the deepfake landscape since 2018. The recent deepfake crisis on X/Grok has woken up millions to the ease and prevalence of AI tools that can create abusive content. However, I worry this will soon be seen as 'simplistic' compared to what it fast approaching, as I shared with Matt: “It’s no longer a very crude synthetic strip,” says Henry Ajder, a deepfake expert who has tracked the technology for more than half a decade. “We’re talking about a much higher degree of realism of what's actually generated, but also a much broader range of functionality.” This functionality, whether it's body pose manipulation, near instantaneous voice cloning, or video avatar generation, are increasingly being fine tuned for sexualised contexts via accessible formats. The harm caused by Grok wasn't particularly novel from a technical perspective, but made the tools of abuse radically accessible and the outputs visible to perpetrators, victims, and digital bystanders. My concern is it's only a matter of time before far more powerful tools and visceral generated content find a similar vessel to Grok/X that 'breaks the surface'. As usual, excellent reporting from Matt Burgess and great to see valuable insights featured from Santiago L., Bruna Martins dos Santos, Asher Flynn, Dr. Pani Farvid, and Stephen Casper.

Definitely one of the more impressive venues where I've delivered a keynote! A real pleasure yesterday to give the keynote address at Trustpilot's Trust In The Age Of AI Summit from the very top of London's Iconic The Gherkin building. Authenticity, transparency, and trust are the foundations of thriving in the AI age, so it was great to provide a map of the frontier through this lens to a packed room, followed by a sky high fireside chat with Trustpilot CEO Adrian Blair. It was also fascinating to hear from Adrian and Alicia Skubick on just how dramatically search has changed in just the last couple of years and how Trustpilot reviews are helping to ground the new wave of "AI answer engines". Big thanks to Stephanie Gillies, Kate Delaney, Jean-Baptiste Gehringer, and the whole Trustpilot team for putting together a brilliant event, it was certainly one i'll remember well!

Disturbing, but sadly not surprising. This finding makes it painfully clear how far we still have to go in the fight against deepfake abuse. A survey commissioned by UK Police found one in four people think there is nothing wrong with, or are neutral about, creating and sharing sexual deepfakes without consent. The survey had a relatively small sample size (1700), but it reflects a trend I and others have observed since 2017: Many people still don't think deepfakes cause real harm. As I'm sure most of you reading know, this couldn't be further from the truth. For eight years I have worked with victims whose lives have been turned upside by deepfake abuse and seen the scale of the damage up close. Just the other day, The Guardian published an article talking about the continuing epidemic of deepfakes in schools, detailing young girls vomiting with distress and fear upon finding out they had been targeted. There is rarely a public engagement I undertake where I don't have a concerned parent asking how they can protect their daughter, or someone sharing a story about someone they know being affected. Deepfakes silence, shame, and steal a person's right to self determination and representation. The vast majority of victims are women and girls, but men and boys are also targeted. While this survey's findings may be disheartening, it's worth remembering the UK has been a global leader in criminalising deepfake abuse and ensuring there are material consequences for offenders. That said, until the harmful reality of deepfake abuse is as widely recognised as forms of physical abuse, we have work to do. Full write-up on the survey in The Guardian: https://lnkd.in/ep4tnkEb

Good to share my thoughts with The New York Times on Denmark's novel approach to combating AI harms by granting people copyright to their likeness. The legislation is being framed as combating deepfakes, but I see it as a much broader and "harm agnostic" approach than many have suggested. As I shared in the piece: "this legislation isn't saying, ‘We’re targeting this specific harm,’” Ajder noted. “It’s saying, ‘This is how we think about identity in the synthetic age.’” In other words, when our synthetic likeness can be hijacked with ease, we need a new framework that at its core protects the right of the individual to control how they appear in an increasingly digital first world. Yes, it may help provide legal recourse for those targeted by deepfakes, but it may also change how companies can train AI on your personal data and our understanding of how photographing or recording someone "IRL" is treated. It's not a flawless approach, particularly with the challenges of enforcement and the worrying fact that copyright is usually transferable, which if applied in this case could lead to people losing control of their own likeness. Nonetheless, it's good to see a country actively attempting to address one of the biggest challenges of AI that is increasingly affecting everyday people. Read the full piece at the gift link below: https://lnkd.in/dKihTvF4

A stark beginning to 2026: I spoke to BBC News about the epidemic of sexualised deepfakes being generated by users of xAI's Grok. The scaling of abuse is shocking: Bloomberg's Cecilia D'Anastasio found that Grok was creating nearly 7000 sexualised or nudifying images per hour (article linked below). When I discovered the first nudification bot on Telegram in 2020, the most striking finding was how integrating nudifying capabilities into a familiar, intuitive platform like Telegram scaled the number of users and volume of deepfakes being created. Now on X, I've personally seen dozens of cases, with targets including children, politicians, academics, and even deceased victims of the recent tragic bar fire in Switzerland. Almost all cases have featured women and girls, with those prompting Grok to create the images seeking to degrade, humiliate, objectify, and silence victims. As I indicate in the featured clip, it's clear the developers didn't intend for it to be used in this way, but this doesn't begin to excuse the release and continued operation of a model that is actively traumatising thousands. It has been good to see regulators in the UK, EU, and India investigating, but i'm not sure if they grasp that this is only the beginning of what's to come on the current trajectory. Powerful multi-modal models, including open source variants, are only becoming more accessible and numerous. At the same time, models that can create sexualised content continue to see significant demand. Without robust model alignment/adversarial testing, thorough legislation, and well resourced enforcement, I worry the continuing collision of these two trends will make this case look trivial for future generations. As usual, excellent work by BBC News' AI Correspondent Marc Cieslak on the report. https://lnkd.in/e8b6xfrb

A little late, but great to see my contributions in the World Economic Forum's excellent new briefing report, Synthetic Data: The New Data Frontier. Synthetic data is a part of the AI stack where I see near universal interest from organisations in both the private and public sector. But it's also an area where there's some of the least grounded discussion, with the actual state of the landscape being obscured by excited projections or ominous prophecies. The World Economic Forum's briefing papers provide much needed clarity on fast evolving areas of AI, and this latest one on synthetic data is no exception. Some key areas the report covers: - Concrete applications of synthetic data already being deployed in areas such as finance, healthcare, and model training. - Evaluations of concerns surrounding synthetic data, including bias, model collapse, and info integrity. - Strategic recommendations for business leaders, developers, and policy makers. If you're looking for a concise and grounded overview of the present synthetic data landscape, capped with actionable insights, I highly recommend giving it a read. As usual, brilliant work by Casey Price, Karla Yee Amezaga, Lauren Woodman, Arun Sundararajan, Cathy Li, Daniel Dobrygowski, Audrey Duet and everyone at WEF/the Global Futures Council who made the report possible.

I've received a lot of requests to comment on OpenAI's new model Sora 2, so yesterday I went on BBC News to share my thoughts with AI Correspondent Marc Cieslak. The TLDR is it's disappointing to see this incredibly powerful tool released in this way. OpenAI did undertake red teaming and have deployed safety features detailed in Sora 2's System Card, namely visual watermarks and C2PA provenance metadata. However, the visual watermarks can be removed (as easily accessible tools have shown) and while C2PA metadata is good to see, there's much more that needs to be done to make it accessible and understandable for everyday people. At the moment, I'd wager most people don't even know it's there, let alone how to access it. But even if these AI disclosure features were fully functioning, it wouldn't negate much of the troubling content being created. In particular, the ease with which people are hijacking other peoples' likenesses without consent due to vulnerabilities in the Sora 2's Cameo authentication processes, is a big concern. Likewise, the volume of highly troubling content being created, even if not technically illegal (open copyright concerns aside), suggests the moderation classifiers need much more fine-tuning. AI videos, often of the 'sloppy' variety, have consistently been some of the most watched content on YouTube and other platforms this year. It's no surprise that OpenAI want to not just drive the means of creation but also distribution. What is a surprise to me is that this product shipped in the way that it did. We're at a critical moment when it comes to generative AI and synthetic media. At the same time that many exciting creative, commercial, and prosocial applications are being released, this year is the first i've started to see the penny drop in wider society about what's at stake if we get this wrong. I imagine that the launch of Sora 2, even if work is already underway to make it more secure, has made that penny drop for many more people.

What are the four horsemen of the 'slopocalpyse' and how should we respond? Watch my recent conversation on AI slop, hosted by the Content Authenticity Initiative, to find out! Bringing together Google DeepMind's Bilva C., Cornell Tech's/ Indicator's Alexios Mantzarlis, and ISD (Institute for Strategic Dialogue)'s Siddharth Venkataramakrishnan, we covered a lot of ground, including: - The rise and many different shades of AI slop, with LOTS of examples. -How AI slop is impacting society and the four key motivating factors driving those creating it. -What role content provenance tech, like Content Credentials from the Coalition for Content Provenance and Authenticity (C2PA)), could play in navigating an increasing slop-filled world. I couldn't have asked for a better group of experts to cover this often disorienting topic- and they were all polite enough to smile at the terrible slop-related pun I was probably saying in this screenshot... Watch the full event and check out the recap here: https://lnkd.in/e_P4nfYw Also don't forget to become a CAI member to join our future expert conversations (we have some exciting ones in the works!), as well as receive invites to exclusive members-only events: https://lnkd.in/gXkGtZ3s

Feeling lucky? When it comes to social media platforms labelling AI generated content, it's still closer to a slot machine than stable digital infrastructure. Alexios Mantzarlis just published his second audit on how reliably social media platforms label AI generated content in his must follow newsletter Indicator. This includes content featuring a variety of provenance signals, including IPTC metadata and Coalition for Content Provenance and Authenticity (C2PA) Content Credentials, as well as proprietary watermarking such as Google's SynthID. Despite Alexios verifying the presence of these signals in the media he tested, the surfacing of them as AI generated content labels was wildly inconsistent (see full results at the link below) Like the last investigation, factors including which AI tool was used to generate the content, the content format (image or video), and how the media was uploaded to the platform (desktop or mobile) had a significant impact, but in ways that were not always obvious or explainable. All the platforms Alexios tested have policies for labelling AI generated content, but none performed reliably, with LinkedIn and Pinterest best overall at 67% of images labelled, while Instagram only managed 14%. So what's the message to take from these findings? As i said in my comments for the investigation, “platforms may be making the right kinds of noises, but the unified front we need to see for the entire media lifecycle is just not there yet.” I'm not looking to trivialise the understandable difficulties in achieving this unified front; we're talking about an incredibly complex patchwork of organisations, pipelines, and media formats. That said, with legislation from the likes of India, South Korea, and California mandating the labelling of AI generated content, don't expect much sympathy from these governments when it comes for enforcement, regardless of whether their demands are currently feasible or realistic. I've not been shy about critiquing approaches to content authenticity that focus on labelling AI noise rather than securing authentic signal, but given the current beleaguered state of the information landscape, it is better than nothing. Yet ultimately, if platforms don't reliably surface the provenance information in an accessible format, it will be just that for the vast majority of users. Be sure to read the full investigation below and subscribe to Indicator. I cannot recommend it highly enough for those looking to understand the modern information ecosystem- both Alexios Mantzarlis and Craig Silverman are killing it! https://lnkd.in/eJUiPjZz

"Criminals are deploying deepfakes because it works. If there’s money to be made and people continue to be fooled, they’ll continue to do it,” This statement may sound obvious, but as I shared with MIT Technology Review's Rhiannon Williams, it's only recently that long running claims of exploding deepfake crime have begun to actually add up. It's not since available toolsets were few and generally behind the SoTA, outputs weren't sufficiently realistic or dynamic, and data or compute requirements/processing speeds made executing attacks very clunky. Granted, one or two high profile cases occurred, but for most cybercriminals, the opportunity cost to experimenting with a janky deepfake was neglecting other proven techniques. Now, as Rhiannon reports and my research over the years has documented, the landscape looks very different. Sophisticated tools that achieve near-parity with SoTA are easily available, AI generated voices and faces are indistinguishable to most, and generation latency is reducing rapidly. The result is the costs for bad actors has changed- deepfakes are no longer a janky, time consuming distraction from tried and tested techniques, but a new proven tool that augments and expands their arsenal. As GetReal Security's Emmanuelle Saliba has covered extensively, deepfakes have transformed the effectiveness of imposter hire attacks targeting some of the world's biggest companies and democratised a wave of fraud, trickling down from the CEO and the celebrity to the everyday citizen. This is far from the endgame for new AI threat capabilities, but it's a good moment to reflect on how we reached this maturation point with deepfakes to better understand how other significant AI threats may emerge and become increasingly viable. Check out Rhiannon's excellent full piece below:

This is a significant release: Cartesia's new Sonic 3 model demonstrates how realism combined with the 'need for speed' is shaping the AI voice race. The expressiveness and realism of Sonic 3 voices across 42 languages are impressive, but the most notable improvement in my view is how quickly it can generate them. Latency for Sonic 3 sits at just 190ms end to end (from input to generated voice output) It doesn't matter how natural a synthetic voice sounds in isolation if there's an unnatural delay in hearing it respond in a natural fluid conversation Likewise, you could potentially over-optimise here and reduce latency to the point that responses are too quick, but natural speech does involve moments of unintended interruption, talking over each other, or interjection. The most realistic voice models will achieve that balance in the right conversational contexts, but it still requires the lowest latency possible- something Sonic 3 is now a leader in. While technically impressive and another step towards broad commercial viability, concerns about misuse are inevitable. The suggestion in Cartesia's launch video that you could use your voice clone to call your friends is clearly intended to be playful, but I worry feeds into the growing distrust that has spread as AI generated content appears more frequently in daily life 'hidden in plain sight'. This isn't to demonise clearly disclosed and thoughtfully deployed synthetic voices. It's reflecting that as voice models grow increasingly powerful, the more urgent questions about security and desirable social norms in the synthetic age become.

It was great to join Clare Duffy on CNN's Terms of Service podcast for this conversation on AI slop, deepfakes, and how to navigate our synthetic world. We covered a lot of ground in just under 30m, including: -The turbulent launch of OpenAI's Sora 2 and the challenge of balancing healthy scepticism towards digital media with the growing trend of AI induced "reality apathy". - The rise and ethical challenges of governing "Techromancy" where AI is being used to 'resurrect' deceased celebrities and family members. - How much of the current frenzy around AI generated content is driven by hype versus substance and what it say about our changing relationship to authenticity. Big thanks to Clare and the team for having me on- I highly recommend subscribing to Terms of Service and checking out their past episodes if you haven't already! https://lnkd.in/eYZ7BWkc

Excited to join Google's programming for SXSW London 2026, alongside Jigsaw's Beth Goldberg and Google's William Carter. Our proposed workshop, My Agent, My Rules: Aligning AI Autonomy and User Control, aims to examine the critical questions of alignment, privacy, and safety that will define the future of consumer agentic systems. Voting is open at the link below, any and all are appreciated! https://lnkd.in/eAj_GBjA

This is worth watching: X are experimenting with AI labels, following a renewed wave of global legislation cracking down on AI content and deepfakes. The final version is yet to be announced and I always welcome experimentation when building approaches to AI/deepfake disclosure. That said, I have my reservations on this initial approach by X based on the years i've spent advising AI companies, platforms, and governments on AI disclosure and transparency: 🏷️ Without a change to X's platform policy requiring users to label AI generated content, there would be no (on platform) consequences to ignoring this toggle. Even if policy were to change, giving it teeth with consistent enforcement is a real challenge for any platform. 🏷️ X aren't the first to experiment with giving users the option to voluntarily label their content (TikTok are another), but it can often look like the platform passing the buck onto the user for a structural platform integrity issue. For good faith users, the label may be welcomed. For bad actors, it will be ignored. 🏷️ Voluntary labelling of this kind still requires a platform verification 'backbone' to avoid it being subverted and actually check that content labelled AI generated IS in fact AI generated and vice versa. By allowing the label to simply be a toggle, it could be easily undermined. The label is only worth the trust users put in it as a reliable signal, so this doesn't just mean being able to label AI content as fake, but also to remove AI labels from authentic content. 🏷️ Even if the label were used perfectly by all users, the current language doesn't help users understand HOW AI has been used. Meta had a similar issue with artist backlash to their initial "Made with AI" label; without clear attribution or explainability, vague labels can muddy the water even further. Given X's Grok nudification scandal kickstarted a renewed focus on regulating AI content, it's not a surprise to see them testing out these labels now. Regardless of what you think about X, it's in everyone's best interest that their approach to AI transparency is well executed. I hope there's more to come and this version is an early test, but as it currently stands it leaves a lot to be desired.

Good to speak with POLITICO's Isobel Asher Hamilton for this Q&A on my work and thoughts on UK AI regulation in today's Morning Tech Newsletter. I firmly believe the UK can become the global leader in creating blueprints for more resilient societies that can confidently navigate the AI age. It's no secret the UK can't compete directly with the US and China on infrastructure, ecosystem, and investment. This isn't to say the UK can't be one of the leading countries in these areas, but it's very unlikely to top the podium. However, the UK is well positioned to become the country the world and industry turns to for guidance in getting AI development and policy right, both commercially and societally. The recent appointment of Kanishka Narayan MP as Minister for AI and Online Safety, as well as leading work coming out of the AI Security Institute, Department for Science, Innovation and Technology, and Ofcom, are smart steps by the government to establish this global influence. I expect and hope there will be more to come in due course!

How good are social media platforms at labelling AI generated content? This essential reality check from Indicator provides the sobering answer. In a first of its kind audit, Alexios Mantzarlis tested AI images and videos across Instagram, LinkedIn, Pinterest, TikTok, and YouTube, to see how many would be labelled as AI generated. The results? No great. Of the 516 posts containing AI generated content, just over 30% received labels, with some platforms failing to label a single piece of content. This is despite all platforms tested stating some form of support for AI content labelling, including the surfacing of C2PA metadata and deployment of classifiers to identify AI generated content. Malicious deepfakes, AI slop, and subtle synthetic media are now ever-present online, making AI transparency and authenticity critical for platforms to get right. As I said to Alexios for the investigation: "this isn’t just about fighting disinformation; this is about providing a whole new infrastructure of trust for the synthetic age.” There are promising signs of this infrastructure gradually developing, but we need to see more stakeholders engaging and adopting. With legislation mandating AI disclosure from the likes of the EU, India, and California soon coming online, I imagine companies and organisations may find urgency in legal necessity... Read the full investigation and see the data breakdown by subscribing to Indicator- it is one of a small number of newsletters I actually pay for- I cannot recommend it highly enough. https://lnkd.in/e7w5G8js

Enjoyed sharing my thoughts with BBC News' Tech Editor Zoe Kleinman on Synthesia's new voice cloning model trained to copy UK regional accents. As AI adoption grows, regional accents in the UK are in decline. AI isn't the primary culprit for this decline, but voice cloning tools that can't accurately recreate certain voices make them less accessible and representative for many groups in the UK and around the world. Synthesia's EXPRESS Voice is a good example that shows improving AI tools isn't just about improving on narrow benchmarks, but providing broader utility to more users. As I said to Zoe, AI doesn't have to "homogenise speech", it can actually help preserve its many nuances and cultural forms. Despite these advances, there are also understandable concerns about the weaponisation of more sophisticated voice cloning. Tools like EXPRESS Voice feature a number of safety considerations in its design and release, but a determined bad actors may be able to compromise these efforts with time, effort, and resources. Yet this risks missing the greater issue in my view. Yes, we should be pushing for secure tools from AI companies, but why try to misuse a defended system when you can use open source tools that are almost as good with no such defences? The open source landscape for voice cloning has matured rapidly in the last 12 months, with my testing of many open voice cloning tools indicating little or no safety alignment in the implementations that are readily available online. In other words, think like a thief: Even if you could, you're less likely to try and break into the house with bars on the windows and security guards patrolling when the house next door is almost as nice and the front door's wide open... Read Zoe's full piece and learn more about Synthesia's Express Voice below: https://lnkd.in/dDJ7JJ3T https://lnkd.in/dMGU8AVq

A fun Halloween find: Thanks to Wikipedia, future generations can learn how terrifying videos of Will Smith eating spaghetti became an AI benchmark 🎃🍝 https://lnkd.in/ecTBFFiP

A good reminder you're likely no AI Sherlock. Less than 10% can distinguish AI generated from real video, according to Runway's new "Turing Reel" research. Based on their Gen 4.5 video model, Runway showed just over 1,000 people two videos depicting the same scene, one real, one AI generated. Only 9.5% achieved statistically significant accuracy, with overall accuracy at 57.1%, not much better than a coin flip. This suggests synthetic video is catching up to Sarah Barrington's findings on our poor ability to reliably identify AI generated voice audio. I've spoken about synthetic video here extensively, but I cannot overstate how the speed of these advances, both in terms of realism and semantic/audio consistency, have been breathtaking. However, I still frequently hear from people who confidently believe they (or often their kids) are able to identify AI generated content reliably and it's old people or the ill informed that get fooled. This simply isn't the case. We need to stop perpetuating the well intended but ultimately misinformed view that media literacy or digital 'street smarts' are enough to reliably identify AI generated content, now and especially in the future. With powerful video models such as Bytedance's recent Seedance 2.0, releasing regularly, we need to put the emphasis on new digital trust infrastructure for the synthetic age, not our outgunned human senses. That said, it's good to see Runway implementing Coalition for Content Provenance and Authenticity (C2PA) to provide transparent and secure metadata to disclose their tools generated the content. We shouldn't discourage people from wanting to get to the bottom of whether content is AI generated or not, but instead provide new literacy, like the Content Authenticity Initiative, that directs them to look for trusted signals, not fast vanishing deepfakes 'tells'. See how you perform on the Turing Reel below:

Yesterday, I was invited to speak at The Labour Party Conference on how the UK government can create a global blueprint for preserving trust in the AI age. There is no sugarcoating that the rapid proliferation of AI generated synthetic media/deepfakes has led to growing suspicion about the authenticity of all digital media. It's telling that "Grok, is this real?" has become one of 2025's standout phrases, as floods of X users ask the (very unreliable!) platform's chatbot to authenticate images, videos, and other media that they now suspect could be AI generated. The scale of the challenge and it's potentially nihilistic conclusion require a reimagining of digital infrastructure in a way that government can, and in my view, should lead. As legislation from the likes of China, Spain, the UAE, and the EU begin mandating the disclosure of AI generated content, the UK has the opportunity to take a slightly different, and in my view more nuanced approach. This is to actively support and encourage the adoption of open standards that provide transparency about the origins of ALL kinds of media, not just singling out AI generated ones. It was a pleasure to share the stage with DSIT Minister Kanishka Narayan MP, Chair of the Science, Innovation, and Tech Select Committee Chi Onwurah, Emily Darlington MP, and Adobe's Matt Day, with thanks to the New Statesman's Harry Clarke-Ezzidio for moderating.

"But how do we KNOW it's real?" For businesses, governments, journalists, and every day people, this question is becoming a reflex in the generative age. So is AI generated content actually changing how we think about knowledge? Or is it revealing cracks in infrastructure and assumptions that we've collectively relied on for too long? Join me next Tuesday 30th September at 5pm UK/12pm ET for a Content Authenticity Initiative conversation with Demos' Elizabeth Seger and Full Fact's Andrew Dudfield on AI and the future of knowledge- from what's really at stake to how we can respond. RSVP below- looking forward to seeing you all there and answering your questions! https://lnkd.in/eQCvzx63

Good to share my thoughts with The Washington Post about Sam Altman's recent warning on AI voice cloning and deepfake fraud. Altman's statement is both nothing new but also revealing about OpenAI's own position on voice cloning. What Altman is saying is something I and others have been warning about since 2018. In July 2019, Symantec reported on three alleged cases of deepfake voice fraud, and we've seen the sophistication and frequency of attacks significantly increase since. So while it's good to see Altman drawing attention to AI voice fraud or 'vishing', it's not exactly a revelation. However, Altman's comments do reflect why OpenAI's much anticipated voice cloning tool Voice Engine likely still hasn't been publicly released. Altman's concerns likely reflect the well documented challenges with securing voice cloning tools and releasing them safely. It's also why several others, including Microsoft's and their powerful VALL-E3 voice cloning model, are yet to see a public release. But as I shared with Shira Ovide, powerful open source voice cloning tools have fundamentally reshaped the landscape, making high quality voice cloning tools worryingly accessible with no safety measures. This doesn't mean we shouldn't continue to expect OpenAI only release tools that satisfy certain safety measures, but that the days where closed AI companies alone hold the keys to hyperrealistic voice cloning are long gone. Responding to this shift is a huge challenge which I can only see getting worse until AI developers and government take coordinated and well resourced action. Read my full comments and Shira's excellent write-up below: https://lnkd.in/ep-N3Sye

So to the surprise of no-one, OpenAI confirmed ads for some ChatGPT users. I spoke to BBC News' Zoe Kleinman about why it was always a matter of time. Despite growing to a massive 800m users, OpenAI still operated at a loss of ~$8bn over the first six months of 2025 according to the FT. So far, there has been no shortage of investors willing to inject cash for OpenAI to burn on the promised path to profitability. However, growing competition for market share from the likes of Google has cranked up the pressure to show meaningful progress on that path. The AI market may be relatively new, but with only 5% of users actually paying for a ChatGPT subscription, ads are a tried and tested solution for OpenAI monetisation. I'm no fan of ads (who is), but at least the proposed 'banner' style ads aren't being integrated into ChatGPT responses directly. As more well funded but deeply unprofitable AI companies start to feel the heat, I imagine we'll see more adopting advertising strategies in an attempt to close the gap, with many being unable to do so. However, it's important to avoid the trap i've seen some make of inferring that a possible economic bubble surrounding AI means the technology's impact on society is also overstated. Just because the average person may not currently be willing to pay for several monthly subscriptions doesn't mean that the technology is going to fade away or won't continue to fundamentally reshape our lives, for better and for worse. https://lnkd.in/ezFF-9qy

Excited to be advising on the UK Government's Deepfake Detection Challenge. If you want to prove you're a leader in deepfake detection, this is your chance! Running 26-29th January at Microsoft's central London offices, the Deepfake Detection Challenge is the world's leading event putting businesses and researchers' deepfake detection tools to the test. Deepfakes and malicious uses of synthetic media have skyrocketed recently, but one of the sharpest points of contrast in the current AI landscape remains the gulf in investment between AI generation and AI detection. As anyone working in the space will know, building robust benchmarking datasets for deepfake detection is challenging, particularly with the speed of developments and new model releases. There’s also the issues of explainability of results, process integration, and navigating adversarial examples in fast moving contexts. The Deepfake Detection Challenge incorporates all of these elements into a programme of high intensity threat scenarios, based on the state of the art UK Deepfake Detection Dataset. There's no silver bullet to the threats deepfakes pose, but detection will, and is already, playing a critical role. Whether that's for better, rather than for worse, depends on leading detection providers participating in critical initiatives like this challenge. Sign up and get involved at the link below! https://lnkd.in/e28sjjB6 Andrew Tyeloo Zac Ghaffar Kate Shanks Varsha Patel Eloise Heyraud (nee Charig) Nicole Lyons UK Home Office Department for Science, Innovation and Technology The Alan Turing Institute Accelerated Capability Environment (ACE) Microsoft

An essential briefing on AI search: Ofcom are brilliant at grounding critical AI discussions. Their new report on "The Era of Answer Engines" is no exception. The shift from traditional to AI search/overviews has been rapid and represents one of the most consequential measures of AI changing consumer behaviour. As agentic systems and AI browsers also gain momentum, we find the nature of information sharing, knowledge production, and the internet itself in flux. With this backdrop, a nuanced overview of the key behaviours, concerns, and pragmatic safeguarding approaches surrounding AI search is timely! As usual, great work as usual by Benedict Dellot, Doris L., and the whole Ofcom team. I highly recommend checking out their other discussion papers:

Your likeness can be cloned from one photo and seconds of voice audio- it's no longer just a concern for celebrities. So how should we respond? Join me today at 12PM EST/5pm GMT for a Content Authenticity Initiative conversation on proving and protecting identity in the age of hyperrealistic avatars. I'll be joined by leading experts Natalie Monbiot (Virtual Human Economy), Kelsey Farish (pioneering AI and media lawyer), and Erik Passoja (SAG-AFTRA). We'll be covering many different perspectives and ending with audience Q&A, so if you have any burning questions on AI, avatars, and authenticity, this is your chance to get them answered! Register for free below, I hope to see many of you there. https://lnkd.in/e8-Ka_sQ

Ajder compared AI slop to smog during the industrial revolution when pollution controls weren't in place: “It's going to be very hard for people to avoid inhaling" Good to provide this critical caveat for The Associated Press' Kelvin Chan guide on how to reduce AI slop content in your digital life: Total elimination is near impossible. I and many others tend to frame the rise of generative AI as like an industrial revolution for the production of digital media. Like the technological marvels that emerged from Victorian Britain and coated the country in soot, the generative revolution comes with its own pollutants- besides those caused by increased energy demands for training/inference. AI slop has become a ubiquitous part of the digital world, clogging feeds and choking human creators and content consumers. Some platforms have started responding. TikTok and Pinterest have introduced controls for reducing AI content in feed, and YouTube CEO Neal Mohan has stated it is one of his priorities for 2026. These measures are worth pursuing, but can only go so far. Returning to the earlier analogy, air pollution remains a serious global problem. New laws and technologies designed to control and reduce it have helped, but they aren't evenly distributed or capable of totally eliminating the harm. Likewise, we may be able to significantly reduce the amount of AI slop we encounter, but hoping to entirely escape it defies the reality of our synthetically saturated age.

Good to be back on BBC News discussing the significant limitations with xAI's measures to combat widespread deepfake abuse using their Grok AI model. By restricting access to specific Grok features to paying members, there has been no meaningful barrier to harm introduced. A determined bad actor is not going to balk at a few dollars for a subscription, nor are they going to struggle to pen a fake name and provide a burner payment method. But even if a bad actor were to use their real name and payment details, this is, as I say in the interview, a purely reactive response. It places emphasis on (in my opinion dubious) efforts to identify perpetrators of Grok-enabled harm, but doesn't stop the harm from occurring in the first place. To interrogate and align Grok to this end would be much trickier, and likely involve addressing tradeoffs re restricting some of the general categories of 'spicy' content that Grok can generate. I'm glad to see government and public momentum against deepfake abuse continue to build in response to this moment. However, I also have no doubt the high profile nature of this story will have created more negative awareness and driven traffic to the tools used to create this abusive content, as we've seen since 2019. Making it incredibly difficult to generate deepfakes is essential, but so too is making it clear that our society treats these digital harms just as seriously as physical ones. At this critical time where deepfake awareness amongst many might be first forming, I would like to see some form of government PSA reinforcing this position.

Fears that deepfakes will generate a synthetic 'fog of war' have circulated since 2018, but its only recently that we've seen them used extensively and with meaningful impact. Notably, we're also starting to see deepfakes and AI generated content being used at similar rates to 'cheapfakes' such as out of context media or video game footage. It shows the cost benefit analysis of bad actors deploying deepfakes over cheapfakes has fundamentally changed: Outputs are more realistic, tools are more accessible, and the cost of creation/deployment has dropped. GetReal's Hany Farid and CNN's Clare Duffy do a great job here breaking down this fast changing frontier and how we can respond. I highly recommend giving it a listen.

Powerful generative AI tools are becoming more accessible, and #AIslop is flooding our digital spaces. What is it, why are people making it, and what can we do about it? CAI advisor and Latent Space Advisory founder Henry Ajder breaks it down. Read the recap and watch the full event we hosted about AI slop: https://lnkd.in/e_P4nfYw Join the Content Authenticity Initiative: https://lnkd.in/gXkGtZ3s

Think authentication tech is just for fighting deepfake disinformation? Think again. Check out my latest video for the Content Authenticity Initiative, where I argue digital authenticity is a growth engine that will redefine how businesses perform due diligence and customers build trust in businesses.

Headlines about political deepfakes are becoming 'boring'. It's not because it isn't worrying, but it's become the new normal in the AI age. So how should we respond? Last week, Conservative Party MP George Freeman was targeted by a deepfake that depicted him defecting to the populist Reform Party. The deepfaked Freeman announced "the time for half measures is over" and the "Conservative party had lost its way", with the video circulating primarily on Facebook. It was quickly debunked, but deepfakes often leave a lasting impression on audiences that remains long after they learn the content is AI generated. Freeman was understandably upset, reporting the incident to the police and arguing that "regardless of my position as an MP, that should be an offence." The big question though, is what will reporting this incident to the police achieve? As Freeman statement suggests, this isn't to say that it shouldn't achieve something. But outside of explicit deepfakes, current UK law does not provide clear provisions for law enforcement to respond to these cases. I've been briefing ministers and politicians on generative AI, deepfakes, and synthetic media since 2018. Over the last 18 months, there has been a noticeable shift in tone as many have started to know colleagues, friends, or family who have been affected, or they themselves have been targeted. Appetite for action is growing, but what that action should look like still faces key debate: - How it safeguards freedom of speech concerns, particularly concerning satire and critical art. - What resources can meaningfully be attributed to enforcement to give any new legislation 'teeth' as the volume of AI generated content continues to skyrocket. - What should appropriate legal consequence/punishment for creating and sharing deceptive/malicious deepfakes look like. We've seen some 'kneejerk' approaches to political deepfakes, such as placing a publication moratorium in proximity to elections, fail to satisfy these questions. They're not easy to answer and as always a pragmatic reality eats policy ideals breakfast. But as critical (non-explicit) deepfake incidents continue to grow, I only hear calls for action from government getting louder. https://lnkd.in/ews3KZJE

Check out my latest short for the Content Authenticity Initiative: How can we fight the growing scourge of deepfake fraud?

This is a significant moment for creators battling to protect their work in the AI age. I often talk about secure signals like Content Credentials as essential infrastructure for navigating our new synthetic reality. Establishing whether content is real or fake is a big part of this, but equally important is how it confirm who created/captured/owns content and how they wish for that content to be handled. Safe Creative are pioneers by putting Content Credentials at the core of how they protect creator's rights and content integrity. They're the first of what I expect will be an adoption avalanche that reshapes the landscape of copyright and content protection. Watch this space!

RIP OpenAI's Sora. Last year, I went on CNN's Terms of Service podcast with Clare Duffy to discuss its launch. Today, here's a few post mortem reflections: As I mention to Clare, it was obvious from the offset that Sora was costing OpenAI a LOT and it sustainability was always in question. Some estimates put the cost at ~$15m a day. Given OpenAI is still unprofitable and pressure from investors and rivals is growing, this is cash they likely decided they can't afford to continue burning. Yet Sora's launch clearly wasn't motivated by the promise of a quick financial return, but likely several strategic bets on AI video: 1. Securing the attention economy- dominating the growing consumption of AI content formats, driving user acquisition, and leader perception in the AI arms race. 2. Controlling the means of distribution of AI content, not just production- given the prominence and viewership of AI generated content on major social platforms like TikTok and YouTube. 3. Showcasing Open AI's capabilities to B2B customers- despite the predictable copyright fiasco, a Disney partnership did emerge (now dissolved) from Sora's launch as bets were made on the future of AI video as the future of content creation and consumption. Ultimately, a good number of these bets have failed: active user numbers shrank, backlash to AI slop content has grown, rivals have released more capable models (e.g. Bytedance's Seedance), and perhaps most tellingly, the novelty that Sora initially brought to many has worn thin. As OpenAI continue to try to forge a path to profitability, the 'blitz scaling' approach to Sora was always going to be a challenging one to turn into sustained monetisation and lasting impact. I highly doubt this is last we'll see from OpenAI in the video space given Sora's outputs are still fairly impressive, but it's perhaps a sign that the inevitable economic reality and growing cultural backlash are starting to bite for even the biggest AI companies.

This is an essential reality check from GetReal Security's Hany Farid and Emmanuelle Saliba about our ability to spot deepfakes reliably. Far too often I see people confidently claiming there's a 'golden rule' for spotting deepfakes- from a lack of blinking in 2018 to glitches with facial occlusion in 2025. Both became redundant fast. The reality is there are no golden rules with a technology like AI synthesis advancing in such a rapid and distributed manner. Suggesting otherwise is, in my view, actively damaging. We need to move beyond this false comfort and acknowledge that we're no longer able to verify media reliably with our eyes and ears. The sooner we do, the more precious time we can dedicate to reinforcing business and societal infrastructure for the deepfake age. https://lnkd.in/eFt5iHBg

A must watch explainer from GetReal Security's Emmanuelle Saliba on Kling 2.6 Motion Control and the ever changing face (literally!) of deepfake threats.