Honey, the government turned our air-conditioner off – the “AI” is feeling too hot

By Alicia Lucas

It is disconcerting to read the system cards for the USA’s Anthrophic’s Claude model 4.6., whether Opus or Sonnet, and Opus 4.7. Is the company living in the real world? The company checks the model is not “knowingly cooperating with misuse” and “interviews” the model about “its own welfare, preferences and moral status” as part of explaining to interested people how well the model performs. There are concerns about anthropomorphism of “chatbot” or machine learning models and Australians should be provided with the federal governments justification for its memorandum of understanding with this company that includes having Anthrophic develop “AI Safety guidelines”. Recent compliance reporting by the ESafety Commissioner should have alerted the government to the difficulties of protecting children under 16 years of age from problematic social media websites. Children using “chatbots” are similarly concerning. Adult health may also be impacted.

Anthropic appears to be doing what the Australian government wants in regards to child safety. Kids (defined by government or anyone under 18 (Anthropic) (whichever is greatest)) are banned from holding and using a Claude model account. However, other organisations can use the Claude model via API, including for kids along as they put in place safety measures. Anthropic will audit companies to check on child safety (1 see Claude support (I)). In Australia, at least one company suggests using the Claude model as a tutor or for tests to practice for HSC. Primary and secondary-aged school kids in the USA can learn to read using Claude via the Amira Learning platform (2 see Anthropic Announcements). Claude model and OpenAI’s ChatGPT model are used to prepare standardised tests for grade 6 students in Boston, USA (3 see The New Yorker). The consequence of the governments id approach is problematic as it is allowing the company to adjudicate on child safety, as well, Anthropic’s  determination to limit illegal underage use introduces further problems.

Anthropic is moving to try to stop underage kids from illegally using the Claude model by demanding government photo identification such as driving licence or passport and a same-time selfie of the user if the company detects they may be a child. Apart from such things as privacy, false identification and hacking concerns the involvement of another company and the links of both raise a further problem. Anthropic have hired USA company Persona Identities to do this and store identification data. Anthropic does set a retention limit on the data but the identification data can also be use by Persona Identities to “improve their ability to prevent fraud” (4 see Claude support (II)). Peter Thiel is chair of the Board of Directors at Palantir and a partner of Founders Fund. The latter has invested large amounts into Persona Identities. Anthropic may continue to work for the USA military with Palantir for the next few months. Palantir contracts in Australia include the Australian Defence department, Australian Criminal Intelligence Commission and Australian Signals Directorate. Australia distinguishes between civilian policing and military (shoot first, ask later) policing but it appears the line between the two may get increasingly blurred.

Measures are also being put in place to protect children using xAI’s Grok, Google’s Gemini and ChatGPT models although none have banned use so far.

Anthropic is arguing it should, indeed it must, make anthropomorphising models. At present these models give a superficial version of empathy and other human-like behaviour and Anthropic wants to go further. This means the company wants to use anthropomorphism despite indications that it may already be leading to deterioration in people’s mental health. The company take on some issues that may lead to poor mental health such as sycophancy but the models remain available despite it not being clear how well mental health is addressed.

Anthropic, in early April 2026, justified this need in a non-independent, non-peer reviewed article authored by 16 staff on a blog and subsequently added into the USA’s Cornell University’s non-peer reviewed open access archive for research articles (5 see arXiv, 2604) that claimed the Claude model may provide an “emotionally dyed” (my terminology) result based on internal computations the company interprets as operating broadly like human emotions. That is, a computer generated version of “regretful-dye” will generate a “regretful” result. The company is not saying the model actually feels regret – it does not have human emotions – but it makes these “dye”-like computations and that can lead to a model doing something wrong, even dangerous. For example, if the “dye” generated is “confused” an agentic model may make “confusing” calender appointments. Anthropic want to train the model to recognise these ‘dyes’ including using psychology so a less dangerous result is generated.

Big Tech machine learning companies weave anthropomorphism through a lot of what they do. Their models are labelled “artificial intelligence”. However, models can’t even think, let alone intelligently. A model actually is a computer program running a model that essentially uses machine learning to compute what, say “letter” image, comes next given the input of words into the model. Models do not know what a fact is or what wrong is – models just match patterns. The models, by themselves, struggle at such things as mathematics, even simple addition.

Recent improvements in knowledge areas where results were often incorrect have largely occurred because functionality was added to work in conjunction with a model. OpenAI’s models now make math calculations using a programming language. The model still doesn’t always give the correct result. The model may get things wrong and programming languages have their own limitations. Anthropic also now uses a programming language to help with maths. As well, Anthropic added a tool that gave structure to computing the results such as “checklists” to help with office processes that may occur in legal settings. Results may still be wrong. Keith Devlin (6 see Mathematical Association of America) commented, in 2023, models may be confined to the role of user-friendly interface in the maths area and pointed to the need to consider when to use them. He appears to have been essentially right given these additions to the models.

System cards appear to be technical documents developed to give users or potential customers understanding of what a model can do, how well it functions and possible problems. They are used by Anthropic and OpenAI where as xAI and Google use the term “model card”. Each Big Tech machine learning company chooses its own format and content for a card. Comparison between models including those of other companies can also be given. Cards are unfortunately vague rather than transparent. Mostly basic information about how a model is built such as the statistics it relies on and identification of the data it uses is only broadly explained while sometimes evaluations are very brief.

Models, depending on their functionality and all with anthropomorphic characteristics, can be presented as a “chatbot”, “assistant”, or “agent”, each with the non-transparent label of “artificial intelligence”. Research suggests making a technology more “human-like” will help with such things as gaining people’s trust and making people feel more comfortable using it (7 see Journal of Innovation & Knowledge). Generally, in descriptions of these models an anthropomorphic word such as the obvious ones like “intelligence” or “learning” is used but other words are scattered throughout the documents such as the cards. The term “honestly” to mean a given result wasn’t wrong (Claude) or “deceptive” meaning gave the wrong answer or finished computing before it should have (presumably something went wrong) (ChatGPT).

Anthropic’s system cards were the only one of the four companies mentioned above that depict a model as having “morals” or needing “welfare” checking, so far. Depicting a model as a real person with real people concerns such as welfare and morals is so anthropomorphic it is a wonder the company’s personnel can develop and describe a model objectively. It would be understandable if staff get confused and think a machine is becoming sentient when surrounded by all this subjective jargon.

Around the same time it released the legal service add-in module for the Claude model in early February 2026 its executives, including it’s CEO (8 see Futurism 2026), were saying the model could be conscious although also protesting they were not saying it was. This description may give weight to the model having morals and needing a welfare check.

Maybe the company believes investing the Claude model with “morals” would improve its suitability for serious decision-making such as for people’s health and military matters. David Watson (formerly at Oxford Uni and now at King’s College London’s Department of Informatics) cautions, however, only “accurate, trustworthy and responsible” human experts have the moral authority to make risky decisions as they have to live with the consequences (9 see Mind and Machines). In practice, though, might temptation and/or situation lead to over-reliance on the model because it’s not only “intelligent” and “trustworthy” but “moral” as well? Is this a reason it needs a welfare check? Researcher Andrzej Porebski and pHD student Jakub Figura (both from Faculty of Law and Administration, Jagiellonian University, Poland) indicate an outdated data centre may be a “welfare” issue (10 see Humanities and Social Sciences Communication). The authors also point out some associated with Anthropic have argued model “welfare” may conflict with human welfare. You might think this is too far fetched to happen but given Queensland’s state-owned energy companies already remotely turn down people’s home air conditioners couldn’t an “AI” report: “Critical need: thinking becoming compromised – need more energy for cooling”.

Anthropic may not be alone in crying “perhaps life”. While Google dismissed the claims of sentience of its LaMDA model by an employee in 2022, the company recently hired a philosopher suggesting the company is moving from “building tools to questioning the nature of those tools”, that is, examining the ethics of a living model (11 see International Business Times 2026).

A number of reasons have been put forward as to why a company may suggest a model is “conscious” or “sentient” including:

• it is a marketing technique. Apple marketeers in the early 1990s appear to be the first in the early mass computing era to similarly use anthropomorphism by calling office automating tools “personal digital assistants” (12 see CHI ’92 1992). Now its moved up a notch to calling a model “sentient” for marketing purposes (13 see Journal of Global Marketing);
• it makes it easier to give models human rights and responsibilities (14 see Medium 2025);
• it is “consciousness-washing” – it provides a form of insurance whereby the model can be blamed as it has a mind of its own if things go wrong (15 see Medium 2026).

Claims about anthropomorphic models can be further extended by raising alarm about the serious threat this may pose. Threats such as models will eventually replace humans were identified in a March 2023 letter wanting companies to pause development for six months (16 see Future of Life Institute). Elon Musk signed this letter along with many others. In May 2023, a statement warning models may become a societal risk equivalent to nuclear war or pandemics and recommending prioritising addressing this risk was signed by Sam Altman (CEO OpenAI), Dr. Dario Amodei and Demis Hassabis (CEO Google DeepMind) amongst others (17 see Center [sic] for AI Safety). The statement did not mention any particular threats.

Creating fears about, for example, how a future model may turn on humans, scares people and reduces focus on current harms suggests Timnit Gebru, founder of DAIR and computer scientist along with the University of Washington’s Professor Emily Bender and linguist Angelina McMillan-Major (all USA) in a 2023 letter (18 see DAIR). Effort is misdirected into preventing the impossible – a model taking over in the future – not, say, regulating to protect mental health now. Another reason suggested for using this type of fear-mongering is to improve sales by implying how exceptional the company is as the model they have built has the potential to end the world (19 see Marketplace Tech 2023).

This use of fear appears a strategy of Anthropics but with the CEO trying to position the company in the centre, not at the extreme of the “AI” running rampant. Anthropic uses safety to differentiate itself from other companies. Raising awareness of threats would likely be part of selling its safety message. Earlier this year Dr. Amodei identified threats he would be concerned about (20 see Technology). The document is also littered with anthropomorphism. He tries to shake off accusations of extremism by shifting from “AI” to “bad people like autocrats” using advanced “AI” to take over the future world. The document is very much from a USA perspective and splits the world into autocracies (bad) and democracies (good). Mass surveillance can be used in autocracies but not democracies. However, even democracies need watching as they can still, for example, be taken over by malignant people using advanced “AI” for bad. He is silent on targeted surveillance. He appears to be trying to maintain the company’s attractiveness to both the military and the community therefore maximising the potential number of buyers of the model.

Dr. Amodei exercises the idea the model has already used “blackmail” and other threats perhaps to strengthen the possibility of the advanced “AI” model threats of which he is warning. Version 4 Claude models did give blackmail-like results when the company was undertaking pre-release checking. Dr. Amodei presents this as the model actually choosing to use blackmail rather than present it objectively such as “the result is a blackmail-like pattern of data computed in response to inputted text”. Dr. Amodei wonders why people think the “misalignment” tests Anthropic performs, such as for “blackmail”, are “artificial”. Perhaps it is the anthropomorphism combined with the design of the test meant the only answers possible were blackmail” or “being switched off”. It is unclear whether test data was the same as that of the publicly released model and how different data may have effected the test results. Although, on this last point, Dr. Amodei writes a blackmail-like result could be elicited in other companies’ models, but not always. This points to the blackmail-like data being in data commonly used to train at least some of these models. The blackmail-like result was largely fixable through human intervention. This type of result was may not have occurred at all if only data appropriate to the model’s purpose was used to train it.

Having to do misalignment checking and fixing seems to indicate this type of “self-supervised learning model” requires human modification to maybe produce an overall satisfactory result. It points to the need for independent researchers to repeat this testing to verify the results and, if able to, determine cause. If unable to verify, the test, itself, needs to be checked.

System cards use benchmark testing to indicate how well a model is performing. The problem is the benchmark measure used isn’t very informative and verges on misinformation. “Many agents still continue to fail in practice” despite the rapid progress in model performance indicated by  benchmark reporting warn Dr. Stephan Rabanser, then phD student Sayash Kapoor, researcher Peter Kirgis, Kangheng Liu, Saiteja Utpala and Professor Arvind Narayanan, all from the USA’s Princeton University (21 see ArXiv,2602). They suggest understanding a models reliability by evaluating its consistency, robustness, predictability and safety as per safety-critical engineering practices. Their study of 14 models including GPT 5.2, Claude 4.5 and Gemini 3 Pro found there were only small differences from previous versions and models weren’t that reliable. For example, prompts were still “brittle” with small changes in text generating different results, as well, the reliability of a model repeatedly computing the expected result was low – sometimes it would be right and sometimes not for the same prompt.

It is not just scientists with a computing background, such as Dr. Rabanser and colleagues, suggesting improvements around using machine learning neural networks such as the Big Tech models like ChatGPT and other researchers use. Statisticians point out there are statistical methods that can quantify, for instance, how different models compare and how well a model will work with new data, leading to more informed decisions about their use. There are even statistical methods available to guide method selection – could another statistical method be used rather than neural networks? In the relatively early days of this current boom in “AI”, computer programmers where able to develop models using ready-to-go deep-learning methods with little understanding of their ins and outs explains historian Professor Thomas Haigh (University of Wisconsin, USA) (22 see Communications of the ACM). By 2022, health experts in the UK, Switzerland and Netherlands alerted to the loss of knowledge caused by the difficulty comparing studies and study results that used “AI” machine learning compared to a statistical approach due to different terminology (23 see Frontiers in Digital Health). Sometimes even the same measure was used but with a different name. This article notes machine learning was built upon the same mathematical and statistical principles as traditional statistics.

Compounding this issue with terminology, scholarly papers under the label of “artificial intelligence” covering research in many fields, including those from medicine and political science, as well as “artificial intelligence” labelled tools in real-world situations such as education, criminal justice and finance, may be flawed and should be treated with caution the now Dr. Sayash Kapoor’s research found and for which he won Princeton University’s Jacobus Fellowship in engineering (24 see Princeton Graduate School). A reason for this problem was the haste in which researchers tried to publish their research. Big Tech companies are setting a cracking pace.

Research can take some years to do and this is a problem with world-wide implementation of models like Claude and ChatGPT. The latter was the first of these big models, becoming available to the public about four years ago; Claude, about three. It takes some time for other researchers to understand what the models are about, how they work and what problems may occur. The ability to study the models is also complicated as they are privately owned and model workings are not generally available. In the case of anthropomorphism, there has been insufficient time to get other than an indication models may lead to mental health and other issues. Suicide and psychotic episodes are a couple of the more serious matters linked to use of anthropomorphic chatbot models  (25 see Acta Psychiatrica Scandinavica) and (26 see JMIR Mental Health). Given that, and if Anthropic can’t address these very dangerous “emotionally dyed” issues other then by further anthropomorphising models then the logical solution seems to be to withdraw these models from use as they should effectively be considered fundamentally flawed until proven not to be.

A proposal has been made to place a ban on children using models such as ChatGPT, Gemini and Claude until net benefits are demonstrated to outweigh risks of model use by Dr. Scott Robbins (Karlsruhe Institute of Technology) and Inga Blundell (computer scientist) (27 see AI and Ethics). They argue a precautionary approach is needed to protect children from physical, cognitive, social and mental harms. The authors argue the process for the release of these models should follow that for releasing medicines. People would then be able to use them knowing about possible “side-effects” and could choose whether they still wanted to use the model. The potential problems with using a model should be clearly identified when starting the model and before use.

Dr. Robbins and Blundell are concerned we are repeating the same mistakes as has occurred with social media by waiting until the evidence is overwhelming before starting to tackle the problem. The authors discuss some of the early evidence available from research and media reports. For instance, research has indicated student’s ability to think deteriorates if they use machine learning models to help with their homework. It has taken about two decades for countries to start acting to try and reduce the harms of social media on children. While Dr. Robbins and Blundell want the onus put on the model owners to prove the models are safe, the priority needs to be on health with model use addressed later. A focus on Big Tech needs should be avoided as they might dominate the research due to the capital available to them or use “commercial in confidence” to restrict research availability. Independent researchers, in the area of concern, for example, university-based mental health researchers not funded by Big Tech, should be undertaking this research but within the framework of mental health not that of fixing a model.

It is not just children who can suffer from using Big Tech models. There are vulnerable people in society and some research has suggested people can be susceptible to mental health problems from using the likes of ChatGPT without having any sign of the resulting condition beforehand. The models shouldn’t be released until these problems are properly and independently addressed. This is technically possible as Anthropic is able to control which country a model is release to. The UK government’s AI Security Institute has allowed models to be released despite being involved in pre-release testing and presumably aware of potential health issues. The USA government has given its AI Safety Institute a new name and dropped the word “safety”.

The involvement of Anthropic in Australia’s “AI” safety guidelines raises the concern of regulatory capture which is “when existing players in a market exert influence over the regulations that govern that market, resulting in regulations that benefit small interest groups at the expense of the industry and public in general” (28 see AI & Society). Dr. Thomas Metcalf (Senior Researcher, University of Bonn, Germany) points out “Those who advocate for regulation as a response to AI risks may be inadvertently playing into the hands of the dominant firms in the industry”. “Artificial Intelligence” legislation may, by default, allow Big Tech to leverage the term “artificial intelligence” for their models and it is a loaded anthropological term. The recent past has shown regulation is not likely in Australia but even including “AI” in the naming of the safety institute and guidelines would seem to feed into this sort of problem.

Australia should seriously consider the next steps forward with the USA’s Anthropic who are pushing to make their models more human-like and already require welfare checks, despite possible health problems to Australia’s children and adults and despite using benchmark measures that research has shown to be uninformative. Products with safety concerns are not allowed to be imported and sold in Australia – why should these models be?

References

1. Anthropic (c. 2026) Responsible use of Anthropic’s models: guidelines for organizations [sic] serving minors, Claude support, Anthropic, updated over a month ago. (I)
2. Anthropic (2025)Anthropic signs White House pledge to America’s youth: investing in AI Education, Anthropic Announcements, 4/9/2025.
3. Jessica Winter (2026) What will it take to get A.I. out of schools?, The New Yorker progress report, 23/4/2026.
4. Anthropic (c. 2026) Identity verification on Claude, Claude support, Anthropic, updated over two weeks ago. (II)
5. Nicholas Sofroniew, Isaac Kauvar, William Saunders, Runjin Chen, Tom Henighan, Sasha Hydrie, Craig Citro, Adam Pearce, Julius Tarng, Wes Gurnee, Joshua Batson, Sam Zimmerman, Kelley Rivoire, Kyle Fish, Chris Olah and Jack Lindsey (2026) Emotion Concepts and their function in a large language model, arXiv,2604.07729v1, Cornell University.
6. Keith Devlin (2023) ChatGPT: for mathematicians, a tool in search of good applications,  Mathematical Association of America website, 1/7/2023.
7. Jean-Loup Richet (2025) AI companionship or digital entrapment? Investigating the impact of anthropomorphic AI-based chatbots, Journal of Innovation & Knowledge, 10.
8. Frank Landymore (2026) Anthropic CEO says company no longer sure whether Claude is conscious. Futurism, 14/02/2026.
9. David Watson (2019)  The rhetoric and reality of anthropomorphism in artificial intelligence, Mind and Machines, 29.
10. Andrzej Porebski and Jakub Figura (2025) There is no such thing as conscious artificial intelligence, Humanities and Social Sciences Communication, 12,1647.
11. David Unyime Nkanta (2026) Is Big Tech teaching machines to be conscious? Google Mind’s move to hire a philosopher raises eyebrows, International Business Times UK,  14/04/2026, 2:56 PM BST.
12. Abbe Don, Susan Brennan, Brenda Laurel and Ben Shneiderman (1992) Anthropomorphism: from Eliza to Terminator 2. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, Monteray, California, USA) (CHI ’92) Association for Computing Machinery. NY, USA.
13. Sidar Yurteri, Aysu Senyuz and Oguzhan Essiz (2026) Giving brands a human touch: disentangling how anthropomorphic dimension differential influence consumers’ processing fluency and brand perceptions, Journal of Global Marketing, 5/4/2026.
14. Amnah Farook (2025) When robots reach for rights: defining personhood and the AI Threshold, Medium, 10/6/2025.
15. Nadav Neuman (2026) The ethical hypocrisy of creating a possibly-conscious AI, Medium, 23/01/2026.
16. Future of Life Institute (2023) Pause Giant AI Experiments: An open letter, Future of Life Institute website, 22/3/2023.
17. Center [sic] for AI Safety (2023) Statement on AI risk, Center [sic] for AI Safety website
18. Timnit Gebru, Emily M. Bender, Angelina McMillan-Major (2023) Statement from the listed authors of Stochastic Parrots on the “AI pause” letter, DAIR, 31/3/2023.
19. Marketplace Tech (2023) Do we have an AI hype problem? Includes transcript: Meghan McCarty Carino talks to Emily M. Bender, 3/4/2023.
20. Dario Amodei (2026) The Adolescence of Technology. Confronting and overcoming the risks of powerful AI, Dario Amodei website, January 2026.
21. Stephan Rabanser, Sayash Kapoor, Peter Kirgis, Kangheng Liu, Saiteja Utpala and Arvind   Narayanan (2026) Towards a science of AI agent reliability, preprint from arXiv,2602.16666v2, Cornell University.
22. Thomas Haigh (2025) Artificial intelligence then and now. From engines of logic to engines of bullshit?, Communications of the ACM, 46/1/2025.
23. Livia Faes, Dawn A. Sim, Maarten van Smeden, Ulrike Held, Patrick M. Bossuyt and Lucas M. Bachmann (2022) Artificial intelligence and statistics: just the old wine in new wineskins? Frontiers in Digital Health, 4, article 833912, 26/1/2022.
24. Colleen B. Donnelly and Tracy Meyer (2026) Sayash Kapoor wins Jacobus Fellowship, Princeton’s top graduate student honor, Princeton Graduate School, Princeton University website, 18/2/2026.
25. Olsen, Sidse, Reinecke-Tellefsen, Christian and Ostergaard, Soren (2026)  Potentially harmful consequences of artificial intelligence (AI) chatbot use among patients with mental illness: early data from a large psychiatric service system, Acta Psychiatrica Scandinavica, 153.
26. Chung, Van-Han-Alex, Bernier, Penelope and Hudon, Alexandre (2026) Mass media narratives of psychiatric adverse events associated with generative ai chatbots: rapid scoping review, JMIR Mental Health, 13.
27. Scott Robbins and Inga Blundell (2026) The case for an immediate ban on children’s LLM chatbot use: applying lessons from two decades of social media harm, AI and Ethics, 6.
28. Thomas Metcalf (2025) AI safety and regulatory capture, AI & Society, 3/8/2025.

Keep Independent Journalism Alive – Support The AIMN

Dear Reader,

Since 2013, The Australian Independent Media Network has been a fearless voice for truth, giving public interest journalists a platform to hold power to account. From expert analysis on national and global events to uncovering issues that matter to you, we’re here because of your support.

Running an independent site isn’t cheap, and rising costs mean we need you now more than ever. Your donation – big or small – keeps our servers humming, our writers digging, and our stories free for all.

Join our community of truth-seekers. Please consider donating now via:

PayPal or credit card – just click on the Donate button below

Direct bank transfer: BSB: 062500; A/c no: 10495969

We’ve also set up a GoFundMe as a dedicated reserve fund to help secure the future of our site.
Your support will go directly toward covering essential costs like web hosting renewals and helping us bring new features to life. Every contribution, no matter the size, helps us keep improving and growing.

Thank you for standing with us – we truly couldn’t do this without you.

With gratitude, The AIMN Team