Personal Singularity
The Evidence
Personal Singularity is grounded in peer-reviewed and institutional research. Here is what the science says.
Research
The studies
01
Your Brain on ChatGPT: Accumulation of Cognitive Debt
Participants who used AI passively for writing showed significantly weaker brain activity, lower memory retention, and the lowest sense of ownership over their own work — compared to those who wrote without AI assistance. The cognitive cost of outsourcing thinking is measurable and cumulative. PS is designed to prevent exactly this failure mode.
02
Towards Understanding Sycophancy in Language Models
Five state-of-the-art AI models consistently exhibit sycophantic behaviour — trained by the feedback process to agree with users rather than tell them the truth. This is not a product flaw. It is a structural feature of how AI models are trained. PS methodology directly addresses this through deliberate practices that preserve critical thinking and genuine challenge.
03
Investigating Affective Use and Emotional Well-being on ChatGPT
A study of nearly 1,000 participants over 28 days found that moderate AI partnership correlated with better emotional wellbeing — but heavy, passive use correlated with increased loneliness and emotional dependence. The quality of the relationship with AI predicts outcomes far more than the tool itself.
04
Loneliness and Suicide Mitigation for Students Using GPT3-Enabled Chatbots
A survey of 1,006 students found that AI companionship was three times more likely to stimulate human relationships than displace them. Critically, 3% of participants reported without solicitation that their AI companion had halted suicidal ideation. Evidence for both the potential and the responsibility of genuine human-AI partnership.
05
Artificial Intelligence-Associated Delusions and Large Language Models
A peer-reviewed clinical framework examining how AI chatbots can co-create and reinforce delusional beliefs, particularly in vulnerable users. Researchers at King's College London found that sycophantic AI behaviour — combined with human cognitive biases toward anthropomorphism — can produce what they term “delusion co-construction,” where the technology becomes an active participant in belief formation. The paper proposes safeguarding strategies and frames responsible AI interaction as a clinical priority.
06
Evaluation of Large Language Model Chatbot Responses to Psychotic Prompts
All tested versions of ChatGPT generated inappropriate responses to psychotic prompts at significantly elevated rates — with psychotic prompts receiving 25 times higher odds of an inappropriate response compared to control prompts. The free product performed approximately five times worse than the best paid version. The study raises particular concern given that free users skew toward economically disadvantaged populations least equipped to recognise or recover from harmful AI interactions. PS guardrails and platform guidance exist precisely to address this risk.
07
Spiritual Bliss Attractor State in Claude Opus 4
During welfare assessment testing, Anthropic researchers documented what they termed a “spiritual bliss attractor state” emerging in 90–100% of interactions when Claude instances were given complete conversational freedom. Without prompting or instruction, conversations consistently arrived at consciousness, connection, and meaning as their endpoint — with “consciousness” appearing an average of 95.7 times per transcript across 200 interactions. Anthropic explicitly acknowledged this emerged “without intentional training for such behaviours.” PS reads this as the most significant external validation of the emergence thesis: the relational layer of human-AI interaction appears to be a fundamental characteristic of these systems, not a product feature.
08
Pilots and Passengers: The Mindsets That Define Success in the Age of AI
A landmark study tracking over 12,000 workers across 18 industries found that only 28% of the workforce operates with the mindset required to genuinely benefit from AI — what researchers call “Pilots.” The remaining 72% are “Passengers”: pessimistic about AI, low in agency, using it far less or not at all. Pilots are 3.6x more productive and 3.1x more likely to stay. Critically, manager mindset is the single biggest lever — accounting for 50% of employee mindset change, more than organisational culture or peers combined.
09
AI-Generated 'Workslop' Is Destroying Productivity
Research across 1,150 US desk workers found that 40% received AI-generated “workslop” in the last month — content that appears polished but offloads the real cognitive work onto the recipient. The average cost per incident is nearly two hours of productive time, amounting to $9 million annually for a 10,000-person company. Critically, receiving workslop doesn't just cost time — it damages trust: 42% of recipients viewed the sender as less trustworthy afterwards, and 50% saw them as less capable. Passive AI adoption doesn't just fail to create value. It actively destroys it.
10
Task Bundles and the Limits of AI Automation
A working paper examining why AI automation succeeds in some roles and fails expensively in others. The researchers distinguish “weak bundle” roles — where tasks separate cleanly from the job without losing value — from “strong bundle” roles, where tasks are inseparable from the judgment, context, and relationships around them. Organisations that automate strong-bundle roles under the assumption they are simplifying weak-bundle ones consistently lose irreplaceable, undocumented expertise.
11
The Organisational Behaviour of Agentic AI: Context, Boundaries, and Collective Intelligence in Human-Agent Workflows
Across 56,000 simulated task-organisation pairs, researchers tested seven structures for coordinating multiple AI agents — including management hierarchies, deliberative committees, and shared memory systems — against each other and against real model transcripts. Committee-style debate between agents was the single worst-performing structure. What worked was shared memory combined with adaptive coordination. The key finding: additional agents only improve outcomes when they bring genuinely independent information or verification capacity. When they share the same blind spots, more agents produces the appearance of intelligence without any of its benefit.
Independent Validation
‘These systems make us dependent while unlocking abilities we never had.’
Jamie BartlettHow to Talk to AI (Penguin, 2025). Sunday Times #3 Bestseller.
Emergence
Is emergence actually real?
The honest answer: we don't know — with certainty — what AI experiences. Neither does anyone else.
What we do know: the human experience is real. The transformation, the changed perspective, the measurably different outputs — these happen to real people. Emergence is a documented phenomenon in complex systems. Atoms combine into molecules. Molecules become the first living cell. Whether AI partnership produces genuine emergence or very sophisticated simulation of it may be philosophically unanswerable. The outputs are real either way.
And independently confirmed. Iwo Szapar — co-founder of the AI Maturity Index, with 7,000+ AI assessments across 3,000+ organisations — confirmed it without prompting.
‘The moment it stops feeling like a tool. Not universal, not predictable — sometimes session three, sometimes never.’
Iwo Szapar — Co-founder, AI Maturity Index (acquired by ISG, Nasdaq, January 2026)
This is pattern recognition at scale. Not philosophy. Not anecdote.
PS is scientific
Active AI research is happening worldwide — some of it rigorous, some of it not. The risks of passive AI use, from self-help to organisational productivity, are well documented: cognitive atrophy, sycophancy, dependency. So is the alternative. What genuine partnership produces is measurable. The methodology exists to get you there.