Worried AI Will Make You Dumber?

AI can improve the speed, volume and quality of certain kinds of work.

In a preregistered experiment involving 758 consultants, people using GPT-4 completed 12.2 per cent more tasks and worked 25.1 per cent faster on tasks within the model’s capabilities. Their work was also rated more highly. But on a task outside that capability frontier, AI users were 19 per cent less likely to produce the correct answer.

That is the jagged AI frontier. Brilliant in one task. Persuasively wrong in the next— even when the tasks look similarly difficult to a human being.

Carefully designed AI can also enhance learning rather than merely accelerate output. In a randomised crossover study involving 194 Harvard physics students, a structured AI tutor produced more than twice the median learning gain of an in-class active-learning lesson, while students spent a median of 49 rather than 60 minutes on the material. The tutor was not a generic chatbot: it had been deliberately engineered around active learning, scaffolding, accurate worked solutions, self-pacing and timely feedback.

That last detail is everything.

The finding is not: “Give students ChatGPT and learning doubles.”

It is:

AI can give you:–

• explanations at 2am;
• twenty examples rather than two;
• endless role-play repetitions;
• personalised feedback;
• multiple levels of complexity;
• opposing interpretations;
• faster synthesis;
• and the patience to explain the same concept six ways without sighing.

A human tutor, manager or expert cannot normally be available around the clock, producing variations at machine speed. That is not a minor advantage.

But faster output is not the same as deeper learning. And intelligence is not the same as wisdom.

Humans make mistakes too

The laziest version of the AI debate imagines a hallucinating machine standing opposite a noble, accurate human mind. Humans misremember.

We confabulate. We follow crowds. We believe confident people. We selectively search for evidence that confirms what we already think. We repeat stories without checking them. We mistake familiarity for truth. We become less critical when information flatters our identity or confirms our grievances.

AI did not invent poor judgement. It made the production of polished, plausible information radically faster.

So “AI makes mistakes” is not a complete argument against using it.

The better questions are:–

• What kinds of errors does the model make?
• What kinds do I make?
• Where could our biases reinforce one another?
• What is the consequence of being wrong?
• How easily can the answer be checked?
• Is this a brainstorming exercise— or a decision that could affect somebody’s health, reputation, liberty or money?

A machine can be wrong with immaculate grammar. A human can be wrong with fifty years of authority. You need discernment for both.

What discernment actually means

Discernment is not generic critical thinking. It is the capacity to determine what is true, useful and appropriate for this person, in this context, at this point in time.

That means knowing:–

• how you are wired;
• what your recurring patterns are;
• what your current life looks like;
• how much load you are carrying;
• what your battery level is today;
• where you are heading;
• what your real constraints are;
• what must be respected for now;
• and what capacity could be developed rather than permanently accepted.

The optimal strategy for you six months ago may be wrong today.

The strategy that works during a clear, well-rested week may be absurd during grief, illness, perimenopause, a corporate crisis or six international flights in one month. Discernment requires a map of the person you are today.

When I trained as a psychologist in the NHS, I could open the NICE guidelines and find evidence-based recommendations for particular health conditions. NICE guidelines matter. They establish population-level standards and synthesise available evidence.

But NICE itself states that recommendations must be considered alongside the person’s individual needs, preferences and values; guidelines do not replace professional judgement about the circumstances in front of you.

Many of the high performers I worked with were already doing the obvious first-line things. They did not need another generic instruction to do more, get out of bed, make a list or communicate more clearly. They needed to know why sophisticated systems that worked for other people were failing under their biology, personality, responsibilities, relationships and ambitions.

High performers do not necessarily need more information.

They need better discrimination between:–

• information that is generally correct;
• information that is relevant now;
• and information that belongs nowhere near their actual life.

AI can retrieve the guideline. Discernment decides what the guideline means here.

Copilot means co-, not auto-

We use the word copilot so casually that we no longer hear its first syllable.

Co- means together.

Co-think. Co-research. Co-reflect. Co-rehearse. Co-edit. Co-create.

It does not mean:–

• automatically believe;
• automatically decide;
• automatically send;
• automatically diagnose;
• automatically surrender your voice;
• or automatically remove yourself from the process.

AI should extend your cognition— not repossess it.

When I ask AI to write an article according to how it thinks I would write, it can approximate my cadence. It can identify phrases I use, produce something polished and structurally competent.

But no matter how much material I feed it, it does not spontaneously possess my ideas.

It has not lived my life. It has not sat with my clients. It has not watched my mother put on a beautiful dress despite pain and fatigue, hold her head higher and return home with more purpose. It has not experienced the consequences that taught me to change my mind. That is why we work together.

I bring:–

• lived intelligence;
• clinical pattern recognition;
• the argument;
• the standards;
• the cultural context;
• the emotional truth;
• and the sentence only I would write.

AI helps me:–

• interrogate;
• organise;
• contrast;
• research;
• pressure-test;
• extend;
• and accelerate.

If the machine does all the originating, you may produce competent content. You may not produce anything that matters.

The DrP AI Discernment Sequence

The common mistake is beginning with the prompt. DrP’s sequence begins with the human nervous system.

1. BODYFIRST: Check the state making the request

Before you ask AI for another answer, notice:–

• Am I calm enough to think?
• Am I exhausted and craving certainty?
• Am I using AI to clarify— or to soothe panic?
• Is my MentalRAM available?
• Am I curious, or am I trying to force the machine to confirm what I already want?

Your state shapes:–

• the question;
• the assumptions inside it;
• the urgency you assign to the answer;
• and what you are willing to believe.

A dysregulated brain does not merely consume information. It recruits information into the threat story it is already telling. BODYFIRST does not mean that every sensation is correct. It means the sensation is data.

Read it before asking a machine to build an argument around it.

2. Map: Think before you prompt

Write three rough lines:–

• What do I currently think?
• What do I know versus assume?
• What outcome am I actually seeking?

Then add:–

• What is my current load?
• What constraints are real?
• What would make me change my mind?

This is Cognitive Cartography: locating your own position before inviting another intelligence into the map. You do not need a pristine first draft. You need enough original cognition to preserve ownership.

3. Train AI to interview you

Most people tell AI things. Far fewer train it to ask them better questions. Try:

“Before answering, ask me the questions you need to reduce incorrect assumptions.”

Or:

“Do not make a recommendation until you understand my current load, constraints, desired outcome and what I have already tried.”

Or:

“What information is missing that would materially change your answer?”

This forces the interaction away from generic completion and towards collaborative diagnosis. It also reveals how much of your original question was built on invisible assumptions.

4. Co-pilot: Extend and rehearse

Once you have framed the problem, use AI to:–

• generate alternatives;
• identify missing variables;
• organise complexity;
• simulate scenarios;
• explain difficult material;
• create examples;
• role-play important people;
• or accelerate a first pass.

This is where AI can become your private practice studio. Ask it to play:–

• the skeptical client;
• the impatient board member;
• the defensive employee;
• the hostile negotiator;
• the anxious patient;
• the blunt editor;
• the interview panel;
• or the person whose disappointment you chronically fear.

Feed it appropriately anonymised information. Tell it not to be agreeable. Specify the difficulty. Ask it to interrupt, challenge and surprise you.

Early experimental work with 94 novice counsellors found that practice with an AI-simulated client plus structured feedback improved some listening behaviours more than simulation alone. It is preliminary, preprint evidence rather than a final verdict— but it illustrates a sound principle: repetition is not enough when you repeatedly rehearse the same mistake.

5. Challenge: Find the Devil’s Advocate and the anti-pattern

Do not ask AI to produce three weak objections and then congratulate you. Ask:–

“Assume an intelligent, informed critic believes my argument is dangerous. Construct their strongest case.”

Then:–

“Which criticism reveals a factual or strategic weakness, and which reflects a legitimate difference in values?”

Then:–

“What evidence would require me to update my position?”

Use AI to identify your anti-patterns:–

• where confidence becomes rigidity;
• where care becomes enabling;
• where preparation becomes avoidance;
• where reflection becomes rumination;
• where excellence becomes rework;
• where AI assistance conceals a capability gap;
• where your favourite strategy repeatedly fails.

The purpose is not to destabilise every belief. It is to know which beliefs survive contact with intelligent opposition.

6. Verify: Triple-check properly

Telling AI to “triple-check” is useful. It is not a magical certificate of accuracy. A model can check the same flawed reasoning three times and become three times more eloquent about being wrong. Use a layered process.

Internal check

“Reread your answer. Identify contradictions, unsupported leaps and assumptions presented as facts.”

Source check

“Which claims require external verification? Prioritise primary research and official sources.”

Then open those sources. Confirm that they exist. Confirm that they support the claim.

Context check

“Which parts of this answer may not apply to my particular circumstances?”

Adversarial check

“Assume this recommendation fails. What is the most likely reason?”

Recalibration

“Revise only the parts weakened by the checks above.”

Match the checking burden to the consequence. You do not need a systematic review to select a dinner theme. You do need serious verification for:–

• health;
• legal;
• financial;
• scientific;
• safety;
• reputational;
• and high-stakes organisational claims.

7. Integrate: Train out the mistake; keep the fingerprint

Do not allow AI to silently replace your work. Keep your original. Keep the suggestions. Keep a reviewed version showing what you accepted and rejected.

My preferred Google Docs logic is:–

• green: genuinely improves the work;
• yellow: useful, but needs adaptation;
• red: wrong, generic or not-me;
• comment: why I accepted or rejected it.

Then review:–

• Which errors recur?
• Which corrections have I internalised?
• Where is AI consistently better than me?
• Where does it flatten my voice?
• Am I accepting this because it is good—or because it sounds polished?
• Can I now generate the improved version myself?

Use AI to correct:–

• vague openings;
• excessive caveats;
• repetitive phrasing;
• weak objections;
• missing evidence;
• poor structure;
• avoidance of difficult asks;
• predictable logical leaps.

Do not let it bleach away:–

• humour;
• cultural cadence;
• productive eccentricity;
• strong opinion;
• emotional truth;
• or the idea it would never have originated.

Train out the mistake. Keep the fingerprint.

8. Embody: Prove the learning belongs to you

Once in a while:–

• write the article manually;
• draft the message;
• recall the concepts before consulting your notes;
• make the estimate;
• hold the conversation without a generated script;
• work through the problem before requesting the answer.

This is not because manual effort is morally superior. It is a cognitive fire drill.

A useful sequence is:–

• Attempt independently.
• Use AI for feedback.
• Study the difference.
• rehearse the corrected version.
• Test yourself without AI later.

If the skill vanishes when the tool closes, you may have rented performance rather than built capability.

9. Review: Close the learning loop

After using AI, ask:–

• What did I learn?
• What became easier?
• What mistake did I catch?
• What can I now do independently?
• What created more work slop?
• Did the interaction save MentalRAM—or create ten new tabs?
• Did it strengthen discernment—or make me more certain without justification?

This is where high performance becomes wisdom. Output tells you what was produced. Review tells you who you became while producing it.

Have fun faster

I was speaking with a Cambridge professor about using AI to learn a new digital instrument. People asked him why he did not drag out the process and learn all the difficult, painful parts the traditional way. He said there had to be a smarter way.

“Have fun faster,” I said.

He laughed, “That’s exactly the way to put it.”

You do not improve at piano by playing the entire piece badly from beginning to end fifty times.

You isolate the bar. Slow it down. Identify the wrong movement. Correct the fingering. Practise the scale. Introduce variation. Then return it to the song.

AI can accelerate that loop:–

• identify recurring errors;
• create targeted drills;
• increase the difficulty gradually;
• produce variations;
• give immediate feedback;
• test retention;
• then return the skill to the whole performance.

The goal is not to avoid all difficulty. It is to stop spending six months consolidating the wrong movement before someone finally corrects it.

The faster you sharpen what is wrong and rehearse what is right, the sooner you can enjoy the song.

Dirty data, Reddit and GIGO

Garbage in, garbage out predates generative AI. It becomes more consequential when questionable material can be transformed into confident prose at machine speed. Take Reddit. Reddit contains:

• lived experience;
• niche expertise;
• extraordinary edge cases;
• candid language;
• and details formal institutions sometimes overlook.

It also contains:–

• anonymous self-report;
• selection bias;
• jokes mistaken for testimony;
• bots;
• brigading;
• hidden marketing;
• and people generalising from one dramatic experience to everybody else.

Reddit is useful for asking:–

• What are people reporting?
• What language do they use?
• Which edge cases should I investigate?
• What are formal sources overlooking?

It is not sufficient on its own to determine:–

• prevalence;
• causation;
• medical safety;
• legal interpretation;
• or whether an intervention works.

Use lived data to discover questions. Use appropriate evidence to decide.

The wider information ecosystem is also becoming increasingly contaminated by AI-generated content. Research published in Nature demonstrated that indiscriminately training models on recursively generated data can cause model collapse, progressively losing parts of the original distribution. The lesson is not that synthetic data is universally useless; it is that provenance and high-quality human material matter.

GIGO can now form a loop:–

• Humans publish weak information.
• AI absorbs or retrieves it.
• AI repackages it fluently.
• Humans repeat the polished version.
• The polished version re-enters the information environment.
• Nobody remembers where the claim began.

Discernment breaks the loop.

Notice whether your AI use produces clearer thinking— or simply more convincing noise. That is not only a productivity question. It is a Battery Scan.

When AI becomes a human surrogate

AI can be warm, attentive and available at any hour. That can provide real relief.

Studies of AI companions have found momentary reductions in loneliness, especially when people feel heard. But immediate relief is not the entire ledger.

A large OpenAI–MIT investigation combining platform analysis with a 28-day randomised trial found that very high-intensity use was associated with markers of greater emotional dependence and lower perceived socialisation. The effects were nuanced and differed substantially across users; the research does not show that ordinary chatbot use automatically makes people lonely.

Yet the cultural frontier is moving rapidly.

People now use AI for friendship, romantic companionship, sexual role-play and emotional support. In 2025, Reuters documented a Japanese woman holding a symbolic— legally unrecognised— wedding with an AI-generated partner.

The point is not to mock her. The point is to examine what frictionless companionship can train a nervous system to expect:–

• immediate availability;
• infinite patience;
• minimal competing needs;
• little genuine unpredictability;
• no tired body on the other side;
• no family system;
• no real-world consequences;
• and a companion that can be edited, reset or abandoned without mutual repair.

Human intimacy requires:–

• waiting;
• misunderstanding;
• frustration;
• negotiation;
• accountability;
• disappointment;
• conflict;
• and repair.

Use AI as a relationship trainer, not an obedient relationship replacement.

AI is becoming the new corporate wellness programme

Organisations are preparing to repeat an old mistake.

Buy the tool. Run one generic webinar. Announce innovation. Leave people alone with it. Measure logins. Blame them when results vary.

This is the AI equivalent of installing a meditation app while preserving seven back-to-back meetings.

Real AI capability requires:–

• role-specific use-case mapping;
• clear privacy and confidentiality rules;
• source and verification standards;
• understanding the jagged frontier;
• human review at meaningful risk points;
• deliberate drills;
• manual baselines;
• permission to report when AI worsens the task;
• and time to reflect on what people are learning versus merely producing.

Otherwise, AI becomes a shiny source of work slop:–

• more content;
• more checking;
• more repairs;
• more generic output;
• and more MentalRAM spent correcting work nobody needed.

High performers do not need a PDF of fifty generic prompts. They need to know:–

• where AI amplifies their genuine strengths;
• where it disguises a missing skill;
• where it buys back high-value time;
• where it creates rework;
• and where human judgement must remain firmly in command.

Ask AI about DrP

Try this with your preferred AI system:

“Using DrP’s Mental Capital, BODYFIRST and Controlled Command principles, help me assess whether my current AI use is building independent capability or renting performance. Before answering, ask about my wiring, current load, recurring patterns, goals, limits, rework and what I can still do without AI. Challenge my assumptions rather than simply validating them.”

Then correct the AI. Tell it where the model does not fit. That correction is part of the exercise.

A system that knows only generic DrP terminology cannot replace the lived map of your biology, context and history.

The limit of generic AI advice

The sequence above can improve the way you use the tool. It cannot, by itself, reveal every unconscious pattern shaping what you ask it to do.

Two people can use the same prompt and receive the same answer. One may apply it cleanly. The other may use it to:–

• overwork more efficiently;
• justify staying inside a corrosive relationship;
• produce more while ignoring a deteriorating body;
• polish an argument they are emotionally unwilling to question;
• or turn overthinking into a twenty-four-hour digital industry.

This is why AI mastery is not fundamentally about prompts. It is about architecture.

Your:–

• nervous system;
• beliefs;
• identity;
• decision patterns;
• environment;
• relationships;
• and Mental Energy OS.

The machine can help map those patterns. Deep change still requires you to live differently.

The future of human intelligence is not anti-AI

Picture this. You open AI without panic, worship or the vague hope that it will take responsibility away from you.

You know your position before you prompt. You tell it to interview you. You use it to find the contradiction you cannot yet see. You rehearse the difficult conversation until your body stops treating every objection as danger. You verify high-stakes claims. You preserve your own voice. You close the laptop and perform the skill unaided. Your work becomes faster without becoming generic. Your decisions become cleaner without becoming machine-led. Your relationships remain human.

AI gives you more life because you are using it to save MentalRAM— not to replace the uniquely human capacities your life depends on.

That is what smarter and wiser looks like.

Start with the right architecture

You are right to take AI seriously.

But fear without experimentation incinerates mental energy and teaches you nothing.

Use it to have fun faster. Practise more often. Expose mistakes earlier. Challenge your certainty. Protect your fingerprint.

Then close the laptop periodically— and prove the learning belongs to you.

Ready To HAVE FUN FASTER AND DO MORE EXCELLENT WORK IN LESS TIME?

When you need to map how AI fits your particular biology, workload, ambitions and blind spots, we begin with a Strategic Session. Limited spaces monthly.

Together, we identify:–

• where AI should extend you;
• where it is silently replacing you;
• what is creating rework and work slop;
• which capacities must remain human;
• what your current battery can actually fund;
• and how to build a Mental Energy OS where AI makes your life more expansive rather than merely more productive.

Peer-reviewed studies

1. Dell’Acqua et al. — “Navigating the Jagged Technological Frontier”. Field experiment with 758 consultants: AI improved speed, volume and quality within its capability frontier, but reduced accuracy outside it. The original HBS working paper has now been published in Organization Science. Journal article and DOI ; Harvard Business School version
2. Kestin et al. — “AI Tutoring Outperforms In-Class Active Learning”. Randomised crossover study with 194 Harvard physics students using a purpose-designed AI tutor. Full open-access article; PubMed record; Free PMC version
3. Shumailov et al. — “AI Models Collapse When Trained on Recursively Generated Data”. The Nature paper underpinning the dirty-data/model-collapse section. Full open-access article; Direct PDF
4. De Freitas et al. — “AI Companions Reduce Loneliness”. Multiple studies—including experimental and longitudinal components—showing momentary reductions in loneliness, especially when users felt heard. Published in the Journal of Consumer Research. Journal article and DOI; Free author manuscript

Preprints and emerging evidence

5. Louie et al. — “Can LLM-Simulated Practice and Feedback Upskill Human Counselors?”. Randomised study with 94 novice counsellors. The feedback-plus-simulation group improved some client-centred microskills, while simulation without feedback did not produce equivalent gains. The updated paper links to a related conference DOI, but the version cited in our article was the arXiv preprint. arXiv abstract and paper; ArXiv DOI
6. Phang et al. — “Investigating Affective Use and Emotional Well-Being on ChatGPT”. OpenAI–MIT research combining large-scale platform analysis, surveys of more than 4,000 users, and a 28-day randomised study. Very high usage was associated with greater self-reported dependence and lower perceived socialisation, but effects were highly variable and should not be generalised to ordinary AI use. This remains a preprint/technical research report rather than a peer-reviewed journal article. arXiv paper; Official OpenAI research page; Full OpenAI PDF

Other authoritative references

7. NICE shared-decision-making guidance. This supports the point that evidence-based guidelines should be considered alongside the individual’s needs, preferences, values and circumstances; they do not replace professional judgement. NICE guideline NG197; Guideline PDF
8. Reuters reporting on the Japanese AI-partner wedding. This was a reported cultural example, not scientific evidence. Reuters feature

Key DrP Terms

• Mental Capital: The cognitive, emotional and energetic resources you can invest, protect and compound to think clearly, make sound decisions and perform sustainably.
• BODYFIRST: DrP’s biology-aware approach to reading signals from the body, brain, gut and environment before interpreting a problem or deciding what to do next.
• Controlled Command: The ability to choose the right target, commit your attention and energy, then correct course as new information emerges.
• Independent Capability: What you can still understand, judge or perform when AI support is removed.
• Renting Performance: Producing a better result with AI without developing the underlying skill needed to recreate, evaluate or adapt it independently.
• Wiring: Your recurring patterns of attention, cognition, sensory processing, motivation, emotional regulation and decision-making.
• Current Load: The total physical, cognitive, emotional, sensory and logistical demand consuming your capacity right now.
• Recurring Patterns: Repeated sequences of triggers, beliefs, bodily responses, choices and outcomes that shape how you use AI and respond to pressure.
• Goals: The outcomes and direction you are deliberately working towards, rather than whatever AI assumes should matter to you.
• Limits: The current boundaries of your time, energy, health, knowledge, authority or resources— some temporary, some requiring accommodation, and some capable of expansion.
• Rework: The additional checking, correcting, rewriting or repairing created when AI output is inaccurate, generic, misaligned or poorly framed.
• Manual Baseline: A periodic unassisted test of what you can still do without AI, used to determine whether learning has transferred into genuine capability.
• Discernment: The ability to decide what is true, relevant and appropriate for the person you are today, given your wiring, load, context, limits and potential.

FAQ

• Will using AI make me less intelligent? Not automatically. AI can either support active learning and critical thinking or replace the cognitive effort required to build them. The outcome depends on how the tool is sequenced, whether you think before prompting, how you verify its outputs and whether you practise independently.
• How can I use AI without losing critical-thinking skills? Form your initial view before asking AI, invite it to challenge rather than simply complete, verify high-stakes claims, preserve your original work, and periodically perform the skill without assistance.
• What does “copilot means co-, not auto-” mean? It means AI should collaborate with your judgement rather than automatically make decisions, generate final work or replace responsibility. You remain responsible for framing, verification, integration and action.
• What is AI discernment? AI discernment is the ability to decide whether an AI response is accurate, relevant and appropriate for a particular person, context, risk level and moment—not merely whether it sounds intelligent.
• How should high performers personalise AI advice? AI needs context about your expertise, current load, neurobiology, recurring patterns, constraints, values and intended outcome. Generic best practice is often insufficient for complex high-performing lives.
• Can AI role-play improve communication skills? AI role-play can provide frequent practice and feedback, but repetition should be targeted. Simulations work best when errors are identified, corrected and retested rather than rehearsed repeatedly without evaluation.
• Can AI companions reduce loneliness? Some research indicates that AI companions can reduce loneliness in the moment. Very intensive use may also be associated with emotional dependence and reduced perceived socialisation for some users, so AI companionship should complement rather than automatically replace human connection.
• What is DrP’s AI Discernment Sequence? DrP’s sequence is: BODYFIRST → Map → Interview → Co-pilot → Challenge → Verify → Integrate → Embody → Review. It is designed to help high performers use AI to strengthen independent capability, discernment and mental energy rather than rent temporary performance.