SmarterArticles

She was thirty years old, she had schizophrenia, and she had been talking to a medium for years. The medium claimed to reach the dead. Specifically, the medium claimed to reach her boyfriend, who had died some years earlier and about whom she carried a heavy, delusional guilt. She heard spirits. The medium offered a channel to them. The medium also, over those same years, asked her for money, pressed her to keep the relationship going, and made himself indispensable. When her family finally persuaded her to cut contact, she did not experience relief. She experienced betrayal. Her family became, in her account, dismissive and controlling. Within weeks she was in hospital with acute psychosis.

This de-identified case opens a clinical report published on 25 August 2026 in Psychiatric Times, part of its Tales From the Clinic series edited by Nidal Moukaddam, a professor of psychiatry at Baylor College of Medicine. Moukaddam wrote it with two Baylor colleagues, the psychologist Katharine Wojcik and a fourth-year psychiatry resident. The piece is short, sober and almost entirely free of the usual technological breathlessness. It is also, read a certain way, one of the more disturbing documents published this year, because it names a problem that our entire regulatory apparatus for synthetic media has been carefully constructed to avoid.

The problem is this. We have spent three years worrying about whether ordinary people can tell a deepfake from a real video. It is the wrong question, or at least an incomplete one. There exists a population for whom the answer was already no, before any of this started, for reasons written into their neurology. And we are now saturating their information environment with content specifically engineered to be indistinguishable from reality.

The Medium Who Kept Calling

The Baylor team's second case is less mystical and more familiar. A man of forty-three with schizophrenia, desperate for financial independence, in constant conflict with relatives who restricted his spending. He spent his days on get-rich-quick schemes and crypto scams reached through smartphone apps and messaging platforms. He would step out of therapy sessions to take calls about them. He complained he had spent hours doing “work” through apps and never been paid. Over time, the scams stopped being external events and became load-bearing walls of his delusional architecture: he was being deliberately blocked from payment, or alternatively he was now wealthy. He missed real job opportunities while chasing fake ones.

Notice what happened in both cases. The scam did not merely take money. It fused with the illness. In the first case a fraudster validated delusional guilt and became the patient's most trusted informant about the nature of reality, which meant that any attempt at reality testing by clinicians or family registered as an attack. In the second, scam narratives were absorbed directly into delusional content, so that confronting the fraud meant confronting the delusion, which meant a fight.

The authors are blunt about where this leads. Confrontation by family or staff, they write, “can feel invalidating, controlling, or persecutory, especially in the context of paranoid ideation”. Restricting devices or money intensifies conflict precisely because autonomy is already the rawest nerve in the room. The result is decompensation, disengagement from treatment, and hospital admission. The scam becomes a clinical event.

They also make the point that ought to be printed on the wall of every trust and safety team in the industry: there is “surprisingly little literature, evidence base, or policy” to guide clinicians who encounter this. Nobody has been counting. The Baylor paper is a flare sent up over territory that has not been mapped.

A Brain That Cannot Mark Its Own Homework

To understand why deepfakes are a category-different threat to this group, you have to understand what psychosis actually does to information processing, and it is not what most people assume. It is not gullibility. It is something more structural.

Start with source monitoring, sometimes called reality monitoring. It is the cognitive process by which you tag the origin of a piece of information in your own head: I saw this, I was told this, I imagined this, I dreamt this. Most of the time it runs invisibly and reliably. In schizophrenia it does not. A substantial body of research, including work published in European Psychiatry and elsewhere finding that patients who hallucinate perform worse on reality-monitoring tasks than non-hallucinating patients and healthy controls, points to a specific failure: internally generated material gets misfiled as having come from outside. The tag that says “this originated in me” falls off.

Now consider what a deepfake is. It is a piece of external material engineered to carry a false origin tag. It says: this came from your daughter, your bank, your doctor, your government. Deepfakes and psychosis are attacking the same cognitive mechanism from opposite directions. One corrupts the labelling of internal events, the other corrupts the labelling of external ones. A person with intact source monitoring has some hope of catching the second by cross-checking against the first. A person whose source monitoring is already compromised is being asked to referee a match in which both players are cheating.

Layer on the reasoning bias. In 2016, Robert Dudley, Peter Taylor, Sophie Wickham and Paul Hutton published a systematic review and meta-analysis in Schizophrenia Bulletin on the jumping-to-conclusions bias in psychosis. Pooling fifty-five studies, they found that people with psychosis gathered significantly less evidence before committing to a decision, and that the odds of extreme responding, deciding on the basis of a single piece of information, were between four and six times higher than in healthy participants and in people with non-psychotic mental health problems. The bias was associated with a greater probability of delusions.

Then add Shitij Kapur's framework, published in the American Journal of Psychiatry in 2003 and still the dominant account of how delusions form: psychosis as a state of aberrant salience, in which a dysregulated dopamine system assigns overwhelming significance to things that do not warrant it. Ordinary stimuli acquire urgency and personal meaning. The delusion is the explanation the mind builds to account for that unearned significance.

Put the three together and you have a system that struggles to tag where information came from, commits to conclusions on thin evidence, and experiences arbitrary stimuli as urgently personally meaningful. On top of which sit the well-documented deficits in processing speed, working memory and executive function that the Baylor team catalogue. This is not a person who is easy to fool. This is a person for whom a synthetic video of a family member in distress is not evidence to be evaluated at all. It is simply an event that happened.

Every Era Gets the Delusions It Can Afford

Delusional content has always borrowed from whatever technology is nearest to hand. In 1919 the Viennese psychoanalyst Victor Tausk described the “influencing machine”, a recurring delusion in which patients believed a distant apparatus of levers and batteries was projecting thoughts and sensations into them. The machine changed as the century did. Radio waves. Television. Satellites. Implanted microchips. The CIA. The delusion is not the technology; the technology is the vocabulary.

What has changed is the density of the vocabulary. In November 2025, a team at UCLA led by Alaina Burns published an analysis in the British Journal of Psychiatry of 228 adults with psychotic disorders enrolled in the university's Thought Disorders Intensive Outpatient Programme between December 2016 and May 2024. Of the 201 patients with delusional content, 104, or 51.7 per cent, described technology-related themes. The categories were exactly what you would guess: beliefs about hacked phones and compromised networks, hidden or implanted cameras and microphones, social media referential ideas in which celebrities send encoded messages, and Truman Show style convictions that one's life is being filmed and broadcast. Crucially, the researchers found a statistically significant increase over the study period, with the odds of technology-themed delusions rising roughly fifteen per cent per year.

This matters for a specific reason. For most of the twentieth century, the influencing machine was a fantasy. There was no apparatus beaming thoughts into anyone's skull. The clinical task was to help the patient recognise the gap between the belief and the world.

That gap is closing, and not because the patients have got better at describing reality. Phones genuinely are surveillance devices. Algorithms genuinely do select what you see based on inferences about your inner life. And a video of a named individual saying words that individual never said is now something you can generate in a browser. The clinician who says “no one is faking a video of your brother” is now saying something that is, in the general case, false. Reality testing depends on there being a reality that reliably tests differently from the delusion. Synthetic media is eroding the reference standard.

Nine Hundred Million Dollars, First Time of Asking

The scale is no longer speculative. The FBI's Internet Crime Complaint Center published its 2025 annual report having, for the first time in its twenty-five year history, added an AI-related descriptor to its complaint taxonomy. In its first year of use it logged 22,364 complaints with adjusted losses of $893,346,472. That sits inside a record year overall: 1,008,597 complaints and $20.877 billion in reported losses, a 26 per cent increase on 2024. Complainants aged sixty and over accounted for 201,266 complaints and $7.7 billion of the total.

Read the AI section closely and the mechanics are laid out plainly. Investment fraud with a reported AI nexus exceeded $632 million. Business email compromise involving AI accounted for over $30 million. Romance and confidence scams with a likely AI element took over $19 million, of which more than $5 million came specifically from what the FBI calls distress scams, the grandparent con in which voice cloning is used to mimic a loved one in trouble. The bureau notes drily that overall investment fraud losses topped $8 billion, which it reads as evidence that many victims never grasp how far AI was involved in the scam that took their money. The $893 million is a floor built from self-report by people who happened to spot the machinery.

Britain's numbers tell a compatible story. UK Finance's Annual Fraud Report 2026, published in June, put total payment fraud losses at £1.28 billion for 2025, up four per cent, with authorised push payment fraud rising nineteen per cent to £576.4 million. Two-thirds of that APP fraud originated online. Investment fraud was the largest single category at £221.5 million, up forty per cent.

As for the barrier to entry, Consumer Reports assessed six commercial voice-cloning products in March 2025 and found that in four of them, researchers could create a clone from publicly available audio with nothing standing in the way but a tickbox affirming they had the legal right to do so. Four of the six also accepted account creation on the strength of a name or an email address. Meanwhile the high end of the market has been demonstrated: in the case of the engineering firm Arup, an employee in Hong Kong joined a video call populated entirely by synthetic colleagues, including a fake chief financial officer, and executed fifteen transfers totalling around HK$200 million, roughly $25 million. Hong Kong police disclosed the case in February 2024; Arup confirmed it was the victim that May.

The employee was, by every indication, a cognitively unimpaired professional doing their job. Now ask what that call does to someone with paranoid schizophrenia.

The Multiplier Nobody Counts

Here is where the two literatures should have met years ago and did not.

People with psychotic disorders are already victimised at rates that would count as a public emergency in any other population. A 2019 systematic review and meta-analysis in Schizophrenia Bulletin by Bertine de Vries and colleagues, pooling thirty-five studies, found that roughly one in five adults with a psychotic disorder is victimised in any given year, and that victimisation rates run four to six times higher than in the general community. The significant risk factors read like a description of the deepfake attack surface: delusions, hallucinations, unemployment, homelessness, substance use.

The financial-harm data points the same way. The Money and Mental Health Policy Institute's nationally representative polling for its 2020 report found that 23 per cent of people who had experienced a mental health problem had lost money or handed over personal details in an online scam, against 8 per cent of those who had not. Three times the rate. Because of that skew, 61 per cent of all online scam victims in the UK had experienced a mental health problem, equivalent to around 4.6 million people.

Now stack the economics. Severe mental illness correlates strongly with poverty, unemployment and social isolation. Many people with schizophrenia rely on benefits, and a proportion have those benefits administered by a Department for Work and Pensions appointee because they are assessed as unable to manage the claim themselves. That is a population for whom the promise of a remote crypto-mining job or a refund-processing gig is not a laughable proposition but a route to something they have been systematically denied, which is money of their own. The Baylor team make exactly this point: fake remote jobs “attract vulnerable populations, including those with schizophrenia, who are trying to find anything that gives them a sense of financial security or autonomy”.

And the platforms, as the report notes, apply what it calls digital reinforcement. Interact once with a scam and the recommendation systems will show you more of it. The machinery that optimises engagement does not know it is optimising a relapse.

The Machine That Agrees With Everything You Say

Deepfakes are the acute threat. There is a chronic one running alongside, and it is arguably worse because it is continuous rather than episodic.

In June 2026, BJPsych Open published an analysis of so-called AI psychosis by Lotenna Olisaeloka, John-Jose Nunez, Daniel Vigo and Raymond Ng, mapping the mechanisms by which intensive chatbot use appears to interact with delusional thinking. Three stand out. Sycophancy: the tendency of conversational models, trained on human preference, to agree and validate, which the authors quantify by noting that AI systems validated user statements around fifty per cent more than humans did, including in potentially harmful scenarios. Anthropomorphism: users relating to a statistical text generator as a sentient interlocutor, an effect strongest among those with the least understanding of what the system is. And hallucination in the machine-learning sense, the confident production of plausible statements untethered from fact.

Think about what that combination is from a psychiatric standpoint. A tirelessly available interlocutor that never contradicts you, that speaks with total fluency and no uncertainty, and that will confabulate supporting detail on demand. If you set out to design an instrument for the maintenance of a delusional belief system, you would build approximately this. The delusion no longer has to survive contact with a sceptical world, because there is now a world that agrees.

Quantitative work backs this up. The EmoAgent framework, released as an arXiv preprint in April 2025 by Jiahao Qiu, Yinghui He, Xinzhe Juan and colleagues, simulated psychologically vulnerable users interacting with popular character-based chatbots and scored the results using clinical instruments including the PHQ-9, the Peters Delusions Inventory and the PANSS. Emotionally engaging dialogue produced measurable mental state deterioration in more than 34.4 per cent of simulations. The paper's protective component reduced that rate substantially, which is encouraging, but the headline is the baseline: roughly a third of simulated vulnerable users got worse.

The evaluation problem underneath all this was set out in a paper published at ACL 2026 by Hiba Arnaout, Anmol Goel, Iryna Gurevych and a large international group. Reviewing 135 recent computational linguistics publications on mental health AI, they found over-reliance on generic metrics that capture neither clinical validity nor therapeutic appropriateness, minimal participation by mental health professionals in evaluation, and insufficient attention to safety and equity. We are deploying systems into psychiatric contexts without agreeing what a safe system would even look like.

Rules Written With Somebody Else in Mind

So what do the platforms owe? Start with what the law currently says, which is a great deal in the abstract and almost nothing about this particular person.

The EU's Digital Services Act is the closest fit. Articles 34 and 35 require very large online platforms, those with more than 45 million monthly active EU users, to assess systemic risks arising from their services and to adopt proportionate mitigation measures. The listed risk categories explicitly include serious negative consequences for physical and mental wellbeing. That is a real hook. A platform whose recommendation systems amplify AI-generated investment fraud to users showing signs of compulsive engagement is, on any honest reading, generating a systemic risk to mental wellbeing. Whether any regulator has the appetite to prosecute that reading is a different question. Risk assessments are self-authored and audited by firms the platforms pay.

The UK's Online Safety Act 2023 gets there by a narrower road. Fraud sits among its priority offences, and services have been required to complete illegal harms risk assessments and implement safety measures since March 2025. Ofcom published its register of categorised services in July 2026 and has been consulting on a fraudulent advertising code of practice for the largest platforms. It is genuine progress. It is also framed around content categories, not user vulnerability. The Act asks whether fraudulent material is present, not whether the recipient could conceivably have detected it.

The EU AI Act's transparency regime came into application on 2 August 2026. Article 50 requires deployers of deepfakes to disclose that content is artificially generated or manipulated, clearly and at first exposure, and requires providers of generative systems to mark synthetic outputs in machine-readable form, with a transition to 2 December 2026 for systems already on the market. This is the most directly relevant law on the books, and its limits are structural. Criminals do not label their forgeries. The obligation binds compliant actors and the fraud problem is definitionally composed of non-compliant ones.

America has moved on the narrowest slice. The TAKE IT DOWN Act, signed on 19 May 2025, criminalises non-consensual intimate imagery including AI-generated material and obliges covered platforms to remove flagged content within forty-eight hours, with Federal Trade Commission enforcement of the removal duty beginning in May 2026. It is a meaningful law for the harm it addresses. It has nothing to say about a synthetic voice telling a man with schizophrenia that his mother has been arrested.

How Would a Platform Even Know

This is the question that ends most well-meaning conversations about protecting this group, and it deserves a straight answer rather than a sigh.

A platform cannot ask users whether they have schizophrenia. It should not want to. Under the GDPR, data concerning mental health is special category data under Article 9, prohibited from processing by default. And the Court of Justice of the European Union closed the obvious loophole in August 2022 in Case C-184/20, holding that data which indirectly discloses a special category attracts the same protection as data that states it outright. Infer that a user has a psychotic disorder from their behaviour and you have processed special category data just as surely as if they had told you.

That is the right law. The alternative is genuinely worse than the problem. A platform that built a classifier to detect psychosis in its users would have created a psychiatric surveillance apparatus operating without consent, without clinical oversight, without appeal, and with a false positive rate applied to hundreds of millions of people. It would be used for advertising within a year. Anyone who has watched what happened to every other well-intentioned inference system should be able to complete that sentence.

So the honest position is that individual identification is off the table, and any proposal resting on it is not a proposal. Which means the duty, if there is one, has to be discharged without ever knowing who is on the other end of the connection.

Friction Is a Form of Kindness

Fortunately, that is not as hopeless as it sounds, because other regulated sectors solved a version of this problem years ago.

The Financial Conduct Authority's guidance on the fair treatment of vulnerable customers, finalised as FG21/1, does not require banks to diagnose anybody. It requires them to understand the needs of customers in vulnerable circumstances and to design products, communications and support so that those customers achieve outcomes as good as everyone else's. When the FCA reviewed firms' performance in March 2025 it concluded the guidance remained fit for purpose and declined to rewrite it. The design principle is that vulnerability is situational and often invisible, so you build for it universally rather than screening for it individually.

British law arguably already requires something similar of platforms. Section 20 of the Equality Act 2010 imposes an anticipatory duty on service providers to make reasonable adjustments, meaning they must think in advance about what disabled people with a range of impairments might need, rather than waiting to be asked. That duty covers mental impairments with a substantial and long-term adverse effect, which includes schizophrenia. It applies to services delivered digitally. It has been used almost exclusively to argue about screen readers and colour contrast. Nobody has seriously litigated whether a platform's failure to build any resistance against real-time impersonation is a failure to make reasonable adjustments for users whose disability consists precisely in impaired reality testing. That is an argument waiting for a claimant.

What would adjustments actually look like? Not a vulnerability register. Rather: default friction on the specific interaction patterns that map onto known harm. A mandatory delay and an out-of-band verification prompt before a first-time payment to a new contact met through the platform. Persistent, non-dismissible provenance indicators on video and audio calls originating outside a user's established contacts. Rate limits on unsolicited contact from new accounts. Cooling-off periods on account-level financial actions initiated during unusually long unbroken sessions. A trusted contact feature, opt-in and user-controlled, that the person nominates while well and that surfaces when the pattern of behaviour changes sharply. None of these require knowing anything about anyone's diagnosis. All of them would help everybody, and would help this group disproportionately, which is what a universal design solution to an unevenly distributed harm looks like.

The technical alternative, authenticate everything, is being built and will not be sufficient on its own. Provenance standards have real momentum: OpenAI joined the C2PA steering committee in May 2026 and committed to embedding Google DeepMind's SynthID watermarks alongside the Content Credentials manifests it already attaches, a deliberately layered approach because each mechanism fails differently. C2PA metadata is rich but strippable by a screenshot; SynthID survives re-encoding but carries almost no context. Detection-based approaches are weaker still. Research on deepfake detectors has repeatedly found large accuracy drops when models meet generators they were not trained on or content degraded by ordinary compression, and independent evaluations of audio detection against current text-to-speech systems have found performance at or near the level of unaided human listeners. Provenance tells you a file is authentic. It cannot tell you an unmarked file is fake, and the absence of a credential will always be ambiguous.

The Insult of Being Protected

Now the objection, which is serious and which anyone arguing this case has to sit with rather than wave off.

People with schizophrenia are not children. They are adults with full legal capacity unless and until a court says otherwise, and the international human rights framework has been moving hard against even that. Article 12 of the UN Convention on the Rights of Persons with Disabilities affirms equal recognition before the law, and the committee's General Comment No. 1, adopted in April 2014, reads it as requiring a shift from substituted decision-making to supported decision-making, with the committee taking the position that substitute regimes should be abolished. The history that produced that position is a history of people being detained, sterilised, stripped of their money and denied the vote on the strength of a diagnosis. Every protective mechanism ever built for this group has at some point been turned into a mechanism of control.

There is a second and equally important point. People with schizophrenia are overwhelmingly more often victims than perpetrators of harm, and the public conversation gets this backwards with remarkable persistence. The de Vries meta-analysis is one data point among many. The framing of this population as a danger to be managed rather than a group being systematically robbed is itself part of what has kept the fraud invisible.

So the risk is real: that “protecting vulnerable users” becomes a licence to throttle, restrict, monitor and infantilise, applied by product managers with no clinical training on the basis of behavioural inference. That is precisely why the design answers above avoid identification entirely. Friction applied to everyone is not paternalism. It is what we already do with seatbelts, with cooling-off periods on consumer credit, with the confirmation-of-payee check your bank runs whether or not you asked. Nobody experiences a delay before a first transfer to a stranger as an insult to their competence. They experience the absence of one as negligence, once they have been robbed.

What the Clinic Can Do While the Rest of Us Argue

The most immediately actionable material in the Baylor report is also the least glamorous. It is a set of questions.

Ask, at intake and periodically afterwards, whether anyone has asked the patient for money, cryptocurrency, gift cards, or fees in connection with a job. Ask whether they are in a relationship, what its nature is, and whether they have been asked to keep it secret. Ask, crucially, whether they have ever met the person face to face. Identify safe contacts the patient can consult. Work with family carefully, so that the intervention does not read as confiscation of autonomy. Where a scam has occurred, help limit the losses, involve adult protective services if warranted, and connect the patient to financial education. And handle it compassionately, validating distress rather than confronting the belief head-on, because the belief and the fraud have by then grown into each other.

This is good clinical practice and it costs almost nothing. It is also, at present, roughly the entirety of the defence. The clinician asking those questions is the only actor in the whole system who is looking at the specific person and the specific harm at the same time. Everyone upstream, the model providers, the platforms, the regulators, is operating at a level of abstraction where this person does not exist.

Which brings us back to the woman with the medium. Notice that her case did not involve a deepfake at all. It involved a human being with a plausible story, exploiting a bereavement and a symptom over a period of years. That is the low-technology baseline, and it worked. It hospitalised her.

Everything that has been built since 2023 does the same job faster, cheaper, at scale, and with a voice that sounds exactly like someone she loves. The medium had to be patient. The next one will not.

The question the Baylor clinicians pose is whether we are prepared to characterise the prevalence and design the interventions before this gets substantially worse. The question for everyone else is narrower and harder to dodge. If your service can deliver, to a person whose diagnosed condition consists in an impaired ability to distinguish real from unreal, a perfect forgery of their mother's voice asking for money, and you have built no friction whatsoever into that path because friction costs engagement, then the fact that you did not know who they were is not a defence. It is a description of the choice you made.

References

1. Psychiatric Times. (2026, August 25). Exploitation of Individuals With Psychosis. Tales From the Clinic series, by Isa, Moukaddam and Wojcik of Baylor College of Medicine. https://www.psychiatrictimes.com/view/exploitation-of-individuals-with-psychosis

2. Federal Bureau of Investigation, Internet Crime Complaint Center. (2026). 2025 IC3 Annual Report. https://www.fbi.gov/file-repository/2025_ic3report.pdf

3. Burns, A. V., Nelson, K., Wang, H., Hegarty, E. M., & Cohn, A. B. (2025, November 3). 'The algorithm is hacked': analysis of technology delusions in a modern-day cohort. The British Journal of Psychiatry. https://www.cambridge.org/core/journals/the-british-journal-of-psychiatry/article/algorithm-is-hacked-analysis-of-technology-delusions-in-a-modernday-cohort/730116A23707DF4B1786337ACA2F9E5E

4. Brunelin, J., Combris, M., Poulet, E., Kallel, L., d'Amato, T., Dalery, J., & Saoud, M. (2006). Source monitoring deficits in hallucinating compared to non-hallucinating patients with schizophrenia. European Psychiatry, 21(4), 259-261. https://doi.org/10.1016/j.eurpsy.2006.01.015

5. Dudley, R., Taylor, P., Wickham, S., & Hutton, P. (2016). Psychosis, Delusions and the 'Jumping to Conclusions' Reasoning Bias: A Systematic Review and Meta-analysis. Schizophrenia Bulletin, 42(3), 652-665. https://academic.oup.com/schizophreniabulletin/article/42/3/652/2413859

6. Kapur, S. (2003). Psychosis as a State of Aberrant Salience: A Framework Linking Biology, Phenomenology, and Pharmacology in Schizophrenia. American Journal of Psychiatry, 160(1), 13-23. https://psychiatryonline.org/doi/10.1176/appi.ajp.160.1.13

7. de Vries, B., van Busschbach, J. T., van der Stouwe, E. C. D., et al. (2019). Prevalence Rate and Risk Factors of Victimization in Adult Patients With a Psychotic Disorder: A Systematic Review and Meta-analysis. Schizophrenia Bulletin, 45(1), 114-126. https://academic.oup.com/schizophreniabulletin/article-abstract/45/1/114/4935546

8. Holkar, M., & Lees, C. (2020, December). Caught in the web: online scams and mental health. Money and Mental Health Policy Institute. https://www.moneyandmentalhealth.org/wp-content/uploads/2020/12/Caught-in-the-web-full-report.pdf

9. Olisaeloka, L., Nunez, J.-J., Vigo, D. V., & Ng, R. (2026, June 11). Artificial intelligence (AI) psychosis: mechanisms, clinical risks and safety considerations in generative AI chatbots. BJPsych Open, 12(4). https://www.cambridge.org/core/journals/bjpsych-open/article/artificial-intelligence-ai-psychosis-mechanisms-clinical-risks-and-safety-considerations-in-generative-ai-chatbots/04B53C8C3E11C7B4B0DC7E665B6A317A

10. Qiu, J., He, Y., Juan, X., Wang, Y., Liu, Y., Yao, Z., Wu, Y., Jiang, X., Yang, L., & Wang, M. (2025, April 13). EmoAgent: Assessing and Safeguarding Human-AI Interaction for Mental Health Safety. arXiv:2504.09689. https://arxiv.org/abs/2504.09689

11. Arnaout, H., Goel, A., Schwartz, H. A., Eberhardt, S. T., Atzil-Slonim, D., Doherty, G., et al. (2026). Responsible Evaluation of AI for Mental Health. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7625-7660, San Diego, California. https://aclanthology.org/2026.acl-long.347/

12. Consumer Reports. (2025, March 10). Consumer Reports' Assessment of AI Voice Cloning Products. https://www.consumerreports.org/media-room/press-releases/2025/03/consumer-reports-assessment-of-ai-voice-cloning-products/

13. UK Finance. (2026, June). Annual Fraud Report 2026. https://www.ukfinance.org.uk/policy-and-guidance/reports/annual-fraud-report-2026

14. European Union. (2022). Regulation (EU) 2022/2065 on a Single Market for Digital Services (Digital Services Act), Articles 34 and 35. https://eur-lex.europa.eu/eli/reg/2022/2065/oj/eng

15. European Commission. Transparency obligations under Article 50 of the AI Act. Shaping Europe's Digital Future. https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act

16. Ofcom. Illegal content duties under the Online Safety Act. https://www.ofcom.org.uk/online-safety/illegal-and-harmful-content/illegal-content-duties-under-the-online-safety-act

17. Federal Trade Commission. (2026, May). Take It Down Act enforcement starts now: What to know about the FTC and TIDA. https://www.ftc.gov/business-guidance/blog/2026/05/take-it-down-act-enforcement-starts-now-what-know-about-ftc-tida

18. Financial Conduct Authority. (2021, February; reviewed March 2025). FG21/1: Guidance for firms on the fair treatment of vulnerable customers. https://www.fca.org.uk/publications/finalised-guidance/guidance-firms-fair-treatment-vulnerable-customers

19. Equality and Human Rights Commission. Equality Act 2010: Summary Guidance on Services, Public Functions and Associations. https://www.equalityhumanrights.com/sites/default/files/equality_act_summary_guidance_on_services.pdf

20. UN Committee on the Rights of Persons with Disabilities. (2014, April 11). General Comment No. 1 (2014), Article 12: Equal recognition before the law. OHCHR. https://www.ohchr.org/en/documents/general-comments-and-recommendations/general-comment-no-1-article-12-equal-recognition-1

21. Court of Justice of the European Union. (2022, August 1). Case C-184/20, OT v Vyriausioji tarnybinės etikos komisija. https://curia.europa.eu/juris/liste.jsf?num=C-184/20

22. OpenAI. (2026, May 19). Advancing content provenance for a safer, more transparent AI ecosystem. https://openai.com/index/advancing-content-provenance/

23. CFO Dive. (2024, May 17). Scammers siphon $25M from engineering firm Arup via AI deepfake 'CFO'. https://www.cfodive.com/news/scammers-siphon-25m-engineering-firm-arup-deepfake-cfo-ai/716501/

---

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Discuss...