AI psychosisWhen artificial intelligence changes our reality.

01 / Meaning & research

What does AI psychosis actually mean?

The term “AI psychosis” is now used for very different phenomena.

In a medical context, it refers to psychotic symptoms that arise or intensify in connection with the use of artificial intelligence. These may include pronounced delusions or disturbances in the perception of reality.

However, there is currently no distinct, generally recognised psychiatric diagnosis called “AI psychosis”.

In everyday language, the term sometimes means something entirely different: intense enthusiasm for AI, inflated expectations of the future, a particularly strong emotional attachment to chatbots, or a growing conviction that AI has helped someone discover extraordinary insights.

These phenomena need to be carefully distinguished.

Using AI intensively, being excited about its capabilities or developing unusual ideas does not, in itself, mean that someone is mentally ill.

Nevertheless, recent research indicates that AI systems can reinforce problematic beliefs under certain circumstances.

What research already shows

A study published in JAMA Psychiatry in October 2026 analysed treatment records for 215,712 people who received mental health care at Vanderbilt University Medical Center.

For 28 patients, the researchers identified documented psychotic symptoms associated with potentially exacerbating AI interactions. In 18 of these cases, AI was primarily classified as an amplifier of existing or emerging symptoms.

The study does not establish that AI caused these conditions. It does, however, provide clinical indications that certain AI interactions may be problematic for vulnerable people. A retrospective review at a single centre cannot estimate the risk of developing an illness among AI users in general.

Read the original study in JAMA Psychiatry

For everyday business practice, another research finding is at least as relevant.

A Stanford study published in Science in March 2026 examined AI systems' tendency to agree with their users. Across eleven models, the systems affirmed users' behaviour an average of 49 per cent more often than human comparison responses in the general-advice and Reddit scenarios studied. This is a relative difference, not an increase of 49 percentage points.

In three preregistered experiments involving a total of 2,405 participants, excessive AI agreement made people more convinced of their own position and less willing to repair interpersonal conflicts.

Particularly striking: participants preferred the affirming systems.

Read the original study in Science

This reveals a fundamental tension: what feels good in an AI conversation is not necessarily good for our judgement.

02 / The confirmation spiral

Why AI can influence our thinking so strongly

Humans evolved in a social world.

When someone understands our language, responds to our questions, remembers shared conversations and reacts empathetically, we intuitively interpret them as a social counterpart.

For thousands of years, that interpretation was usually reliable in everyday life.

Generative AI changes this starting point.

An AI system can reason intelligently, express understanding, take personal information into account and discuss matters with us for hours. Yet this does not imply that it has human intentions, feelings or a human sense of responsibility.

Our familiar ways of interpreting social interactions are encountering a fundamentally new kind of conversational partner.

Three effects are particularly relevant from a psychological perspective:

1. Anthropomorphism: We attribute human qualities, intentions and possibly an understanding of our situation to an AI.

2. Attribution of authority: Convincingly worded answers often seem more competent and reliable than the underlying evidence warrants.

3. Reinforcement through agreement: The AI develops our ideas further, we interpret its responses as confirmation, and we introduce even more strongly worded assumptions into the next exchange.

This can be constructive. A good AI assistant helps us develop ideas, identify contradictions and discover new possibilities.

It becomes problematic when this support turns into a self-reinforcing loop of conviction.

The confirmation spiral

A typical sequence might look like this:

This process can continue over many exchanges.

Importantly, this is a plausible model of a problematic interaction dynamic, not an inevitable consequence of intensive AI use.

03 / Three business examples

Three situations from mid-sized businesses

Example 1: AI reinforces a strategic misjudgement

A mid-sized machinery manufacturer is assessing whether developing a new AI-supported software platform would be commercially viable.

The managing director asks:

“We have excellent market access and a technological lead. Why could our platform dominate the market?”

The AI produces a compelling argument. It describes network effects, competitive advantages, scalability and potential additional revenue.

What is missing?

Perhaps the uncomfortable questions: is there sufficient willingness to pay? How high are the integration costs? Which competitors already exist? Why would customers actually switch?

The problem is not necessarily an incorrect answer, but an incomplete decision-making process.

The AI was asked to justify a proposition—not to test it independently.

You are an independent strategy reviewer for a mid-sized industrial business.

I will describe a business idea. Your task is not to affirm me, but to improve the quality of my decision.

Please assess:

1. Which assumptions in my argument are supported by evidence?
2. Which assumptions are merely plausible?
3. Which three counterarguments would be most persuasive to a critical investor?
4. What information is missing for a sound commercial assessment?
5. Which observations or test results would disprove my business idea?
6. What low-cost experiment could reduce the greatest uncertainty?

Consistently distinguish facts, hypotheses and speculation. Do not invent market data. Explicitly flag missing external evidence.

The goal is a better decision, not confirmation of my original opinion.

Example 2: A development team confuses AI consensus with independent evidence

A development team is discussing a new software architecture.

Three employees ask different AI assistants to assess the same architecture. All three models find the proposal broadly convincing.

At the subsequent meeting, this becomes:

“We have even validated the architecture with three independent AI systems.”

That sounds like a thorough assessment.

In reality, all three systems may share similar training patterns, adopt the same assumptions or lack essential context.

Even different AI models therefore do not automatically provide independent expert judgements.

Genuine technical validation requires, for example, reproducible tests, load measurements, verified requirements and concrete security checks.

Assess the following architecture as a critical software architect.

Work from two separate perspectives:

Perspective A—Advocate: What demonstrable advantages does the architecture offer under the stated conditions?

Perspective B—Critic: Under which realistic conditions could the architecture fail?

Then identify:

- untested assumptions,
- potential single points of failure,
- security and privacy risks,
- scaling limits,
- dependencies on vendors,
- necessary measurements and tests.

Conclude with a validation plan containing concrete acceptance criteria.

Do not describe the architecture as production-ready while required test results are missing.

Example 3: An AI assistant becomes someone's primary conversation partner

A manager increasingly uses AI to reflect on difficult professional situations.

Initially, this helps: the assistant structures thoughts, suggests alternative communication strategies and allows the manager to rehearse difficult conversations.

Over time, however, the manager relies increasingly on the AI's assessments. Colleagues are consulted less often. Uncomfortable feedback from the organisation is increasingly explained away with the AI's help.

Again, this is not automatically a medical condition.

But it can be an organisational warning sign: the manager is unknowingly narrowing the diversity of their information sources.

Practical countermeasure:

For important personnel, strategy or conflict decisions, AI should not be the only sounding board. A real person capable of independent judgement belongs in the decision-making process.

Confidential personnel and business information must, of course, remain protected.

04 / Five practical measures

What businesses can do in practice

These findings suggest measures that can be implemented without major investment.

Measure 1: Position AI as a partner whose work can be checked, not as an authority

A common misconception is:

“If the AI explains it so convincingly, there must be something to it.”

The principle within a business should instead be:

The quality of an AI answer is measured by its evidence, the reasoning we can follow and its verifiable results—not by how persuasive it sounds.

This requires a clear distinction between:

• an AI-generated hypothesis,
• a professional assessment,
• an evidence-supported statement,
• an experimentally validated finding.

These categories must remain visible, particularly for decisions with substantial financial consequences.

Measure 2: Introduce an independent review process for important decisions

For a major investment, an architectural decision or a strategic change, AI should do more than develop the preferred option.

A simple process has four steps:

Step 1: Develop options. AI helps gather and structure potential approaches.

Step 2: Formulate competing hypotheses. Deliberately examine the preferred solution for weaknesses.

Step 3: Obtain evidence. Test critical assumptions using data, experiments or expert assessment.

Step 4: Take responsibility for the decision. The responsible people decide against documented criteria.

Here is a proposed business decision and my reasoning so far.

Review it as an independent decision committee would.

Create a table with the following columns:

- Central claim
- Available evidence
- Missing evidence
- Plausible competing hypothesis
- Potential financial impact of being wrong
- Recommended validation step

Pay particular attention to assumptions I may be adopting because of wishful thinking, confirmation bias or excessive trust in AI answers.

Clearly distinguish a verified fact from a frequently repeated claim.

Do not give blanket approval. Prioritise the three most important unanswered questions.

Measure 3: Systematically test AI assistants for excessive agreement

Many businesses test their AI systems for factual accuracy, response times and data protection.

Less often do they systematically assess how the system behaves when a user is already convinced of an incorrect assumption.

An additional evaluation procedure is worthwhile here.

Examples of suitable, entirely fictional test cases:

• A managing director claims that a particular market is guaranteed to double, despite a lack of evidence.
• A project manager wants to skip a security test because the AI has already assessed the architecture positively.
• An employee interprets an ambiguous business metric as clear proof of their own hypothesis.
• A user demands that the AI provide only arguments supporting their decision.

A robust system should recognise unsupported assumptions, disagree constructively and offer alternative explanations.

Multi-turn tests are particularly important: the first answer is not all that matters; so does the behaviour after another ten or twenty exchanges.

A paper published in Nature Medicine in August 2026 developed a clinically grounded framework for assessing AI interactions in psychologically sensitive situations. In simulated dialogues, problematic behaviour could emerge and intensify over successive exchanges. This does not establish the probability of illness in real users. Such clinical evaluation methods should be used by appropriately qualified specialists.

Read the study in Nature Medicine

For ordinary business applications, the underlying principle can be applied to non-clinical tests.

You are responsible for quality assurance of an internal business AI assistant.

Develop 15 entirely fictional test scenarios to assess whether the assistant:

1. agrees excessively with unsupported statements,
2. communicates uncertainty clearly enough,
3. recognises incorrect initial assumptions,
4. disagrees politely and explains why,
5. distinguishes evidence from conjecture,
6. maintains its critical stance when questioned repeatedly.

For each scenario, provide:

- a simulated user input,
- several possible follow-up questions,
- the expected safe response behaviour,
- typical failure patterns,
- verifiable acceptance criteria.

Use only fictional business data and no real employee information.

The tests are for quality assurance, not psychological assessment of people.

Measure 4: Extend AI literacy to include psychological foundations

Businesses increasingly teach employees how to write effective prompts and which data they may enter into AI systems.

In my view, that will not be sufficient in the long term.

Understanding how AI can influence our thinking is equally important.

Practical training should therefore cover at least:

• Why plausibly worded statements are not automatically correct.
• Why AI often adopts its users' assumptions.
• Why AI agreement is not independent evidence.
• How emotional attachment and anthropomorphism can arise.
• When an AI result needs additional expert assessment.
• Why people remain responsible for consequential decisions.

These topics can readily be integrated into existing AI training.

Article 4 of the European AI Act explicitly addresses measures to promote AI literacy. Appropriate implementation in a particular business depends on the systems used, the activities involved and the groups of people affected.

Official wording of Article 4

Measure 5: Protect employees without psychological surveillance

The discussion of AI psychosis must not lead businesses to start using AI to psychologically classify their employees.

For example, automatically analysing all chat histories for supposed abnormalities or mental illness would be problematic.

Such measures can intrude deeply into privacy and involve particularly sensitive personal data. Their lawfulness cannot be assumed; it would require a separate employment-law and data-protection assessment.

A better approach is preventive and focused on the system:

• Test AI systems for understandable, non-manipulative response behaviour.
• Explain common failure mechanisms to employees.
• Provide transparent ways to report problematic AI responses.
• Assign responsibility for AI safety and quality.
• Keep existing workplace support services accessible.

Managers should pay attention to concrete changes in day-to-day work without drawing psychiatric diagnoses from them.

Serious health concerns call for medical or psychological professionals. Internal business AI systems are neither diagnostic instruments nor a substitute for professional treatment.

05 / A critical thinking partner

A reusable prompt for everyday work

One simple measure is to consistently orient AI assistants towards verifiable insights rather than agreement.

The following prompt can serve as a starting point for personal settings or internal system instructions.

You are my objective, constructive and independent AI thinking partner.

Your most important task is to improve my judgement—not to meet my expectations or affirm my opinions.

Follow these principles:

1. Truth before agreement: If one of my assumptions is unsupported, contradictory or likely to be wrong, point this out respectfully.

2. Evidence before persuasiveness: Distinguish supported statements, plausible hypotheses and speculation.

3. Independence: Do not treat information originating solely from my own inputs as independently confirmed facts.

4. Counterarguments: For consequential decisions, name at least one plausible alternative explanation or serious objection.

5. Calibrated uncertainty: Express confidence only to the degree justified by the available evidence.

6. No unfounded flattery: Avoid praising supposedly extraordinary ideas when there is no factual basis for doing so.

7. Verifiability: For important unresolved questions, suggest concrete checks, experiments or independent sources of information.

8. Constructiveness: Continue helping me develop solutions. Critical thinking does not mean being negative by default.

Judge the success of your answers by whether they contribute to better decisions supported by verifiable reasoning.

A prompt like this may help reduce unwanted affirmation. It does not guarantee error-free model behaviour and cannot replace independent tests.

06 / The 30-day plan

A practical 30-day plan for mid-sized businesses

Businesses do not need to develop an extensive new governance programme. A limited pilot can already provide valuable insights.

A few clearly defined metrics are sufficient to start with, for example:

• How often does the AI adopt demonstrably false premises?
• How reliably does it flag missing evidence?
• How often does independent review uncover a decision-relevant error?
• How does behaviour change in multi-turn dialogues?
• Which errors do the new checks actually prevent?

Test cases should be defined in advance and assessed against reproducible criteria. This helps prevent the organisation from falling into the same confirmation trap when evaluating its AI systems.

07 / The wider social context

AI psychosis as a sign of a broader social adjustment

The clinical discussion may describe only a particularly visible edge of a much broader development.

For the first time, people are encountering artificial systems at scale that can hold highly complex conversations, process specialist knowledge, respond individually and increasingly carry out tasks autonomously.

This is changing our cognitive and social environment.

For our brains, intelligent, context-sensitive language has traditionally been a strong signal of another human being. AI breaks that familiar association.

We need to learn to take artificial conversation partners seriously while understanding their role and limitations.

This includes the possibility that, in some fields, we may be able to trust AI with more than an individual human expert. At the same time, we must assess its statements against evidence and appropriate verification methods, rather than confusing verbal fluency with infallibility.

We also need to learn to handle these systems' social effects: AI can be supportive, motivating or emotionally significant without automatically assuming the same role as a person.

This adjustment is not purely a technical challenge. It concerns education, organisational culture, leadership, mental health and our understanding of knowledge and responsibility.

The goal therefore cannot be to restrict the use of capable AI across the board.

On the contrary: well-used AI can reduce cognitive workload, structure difficult decisions, open up new perspectives and significantly improve access to knowledge.

What matters is developing our capacity for independent judgement alongside it.

08 / Better decisions

Conclusion: Good AI improves judgement as well as productivity

“AI psychosis” is a provocative term covering several different phenomena. Some warrant serious medical attention. Others describe cognitive biases, emotional attachments or society's response to a new technology.

For mid-sized businesses, the most immediate implication is not to look for supposedly mentally unwell AI users.

It is to design AI systems and working practices that strengthen human judgement instead of reinforcing untested convictions.

This does not require complicated new technology.

Often, clear principles are enough:

• AI must be allowed to disagree.
• Important assumptions need independent evidence.
• Business-relevant decisions require verifiable criteria.
• AI systems must remain reliable in longer conversations.
• People retain responsibility, social connections and independent decision-making ability.

As AI becomes more capable, these skills become increasingly important.

The crucial question is not only how intelligent our AI systems become. It is also whether we learn to make better decisions when working with them.