PISA 2026 and AI: Are We Using AI to Learn, or to Avoid Learning?

The latest PISA results arrived at an awkward moment for anyone responsible for learning at work. Generative AI is becoming better at exactly the kinds of tasks that used to require people to read, compare, formulate and explain. At the same time, the OECD is reporting weaker performance in reading, mathematics and science, together with concerns about how students process information.

The study people are searching for as "PISA 2026" is officially PISA 2025, because the assessment took place in 2025. The OECD published the results on 8 September 2026. Around 760,000 students from 91 countries and economies took part.

There is an obvious temptation to turn the results into a simple story about AI making young people worse at thinking. The data does not support that conclusion. Reading performance had been deteriorating before generative AI became widely available, and PISA is an observational study with many factors in play. Motivation, reading habits, classroom conditions, socioeconomic background and broader digital behaviour all matter.

The more interesting part of the report is narrower. PISA gives us another reason to distinguish between using AI to produce a better answer and using AI in a way that helps someone become better at producing answers themselves.

What PISA actually found about AI use

PISA asked students how they used AI chatbots for schoolwork, including for research, summarising, drafting and learning. The relationships with performance were mixed. Students who did not use AI for some specific tasks often scored better than students who did, while moderate use sometimes looked better than either very low or intensive use. The OECD repeatedly cautions that these associations should not be read as causal effects.

One finding is especially useful for people designing learning systems. Students who used AI frequently to help them learn and who also had opportunities to learn how to evaluate AI-generated information tended to perform slightly better than similar frequent users without that guidance. The OECD's recommendation is therefore not to keep AI out of education. It is to use it in targeted ways, including feedback and personalised practice, while preserving the learner's own effort.

That sits alongside another uncomfortable finding. PISA reports deterioration in reading behaviours that matter when information is abundant: evaluating information, connecting material across sources and persisting with difficult text. The proportion of students the OECD describes as "hasty readers", those who move quickly through material and answer incorrectly rather than working through it carefully, rose sharply between 2018 and 2025.

Again, none of this proves that AI caused the decline. It does raise a practical question about the design of learning in an environment where a plausible answer is available almost instantly.

A better result can hide a missing repetition

Consider a buyer preparing for a supplier negotiation.

Ten years ago, preparation might have required an hour of reading market material, sketching the supplier's likely arguments, working through cost drivers and writing down possible responses. Today an AI assistant can do a respectable first pass in minutes. It can summarise the supplier's position, challenge assumptions, propose questions, build a concession plan and draft several responses to an expected price increase.

That is useful. There is no virtue in making someone spend an hour on work a machine can do better in five minutes.

The difficulty appears when the organisation assumes that a better preparation document means the buyer has developed more negotiation capability. Those are different outcomes.

Suppose the supplier rejects the buyer's cost model in the first minute of the call, says another customer will take the capacity, then offers two percentage points of movement in return for an immediate volume commitment. Nothing in the preparation document can make the decision for the buyer cleanly. They have to listen, judge whether the capacity claim is credible, understand the value of the trade, decide what to reveal and formulate a response while the conversation continues.

If AI wrote the arguments beforehand, the buyer had access to better arguments. That does not tell us whether they can retrieve the logic under pressure, adapt it to a change in the situation and say something useful without a prompt in front of them.

This is the part of professional learning that becomes easy to overlook. A tool can improve the quality of today's work while removing one of the repetitions through which tomorrow's capability would otherwise have been built.

Not all effort deserves to be preserved

There is a bad version of this argument that treats difficulty as inherently valuable. It is not.

Nobody becomes a stronger negotiator because they manually copied market data into a spreadsheet. A manager does not build leadership skill by formatting a status report. A lawyer does not improve legal judgment by repeatedly drafting boilerplate that a system can produce accurately.

The relevant question is which part of the effort is actually connected to the capability the organisation wants to preserve.

If the capability is analysing a supplier's cost structure, the person should still have to understand why the analysis works, even if AI accelerates the arithmetic and research. If the capability is handling resistance in a negotiation, reading a suggested response is useful preparation but it cannot replace the experience of producing a response while another person is pushing back. If the capability is giving difficult feedback, a perfectly written email is not equivalent to managing the employee's reaction in the room.

This is where the OECD's emphasis on feedback and personalised practice becomes relevant beyond education. AI is unusually well suited to creating practice conditions without necessarily taking over the part being practised.

Use AI to create the problem, not always to solve it

The same technology can be used in almost opposite ways.

A buyer can ask AI, "What should I say when the supplier refuses my position?" and receive a polished answer. Or the buyer can enter a simulation where the supplier refuses the position and has to decide what to say next.

A manager can ask for the ideal script for a difficult performance conversation. Or the manager can practise the conversation against a counterpart who becomes defensive, asks an unexpected question and challenges the premise of the feedback.

In the first case, AI reduces the amount of work required to produce an answer. In the second, it creates a low-risk opportunity to perform the skill.

That difference is easy to lose in corporate learning because both experiences can be marketed as "AI-powered training". The useful test is simpler: who is doing the difficult part?

If the learner is supposed to become better at making decisions under pressure, they still need moments where the decision is theirs. If they are supposed to become better at speaking, they need to speak. If the skill is adapting to resistance, the other side has to resist.

The AI can make the scenario more realistic, vary the difficulty, provide a counterpart on demand and help analyse what happened afterwards. Those are substantial advantages. They do not require it to remove the learner from the task.

The risk for companies is confusing assistance with capability

This matters because organisations are adopting AI fastest in areas where the output is easy to inspect. A cleaner report, sharper analysis or better email looks like improvement immediately. Capability is harder to see. It becomes visible later, usually when the situation changes and the person can no longer follow the prepared path.

That can produce a false sense of progress. The team appears more capable because its documents are better. In reality, the system around the team may simply have become more capable.

There is nothing wrong with that if productivity is the goal. Companies should automate work that does not need human effort. The mistake is counting the productivity gain as evidence that the underlying human skill has improved as well.

For procurement, that distinction is especially important because the expensive part of a negotiation often happens in the live exchange. The analysis can be excellent and the deal can still move in the wrong direction when the supplier escalates, leaves a silence, introduces a condition or challenges the buyer's authority.

Voice2Evolve was built around that problem. The AI does not negotiate the real supplier deal on behalf of the buyer. It provides conversational sparring for procurement negotiation training: the buyer trains against an AI sparring partner. The buyer still has to listen, decide and speak, and the analysis afterwards is based on what actually happened in the conversation.

The PISA results do not prove that this is the right model for workplace training. They do something more modest and, in my view, more useful: they make it harder to assume that access to better answers automatically produces better learning.

As AI gets better, companies will need to be more deliberate about the difference. Automate the work that deserves to disappear. Keep the repetitions that build the capabilities people will still be expected to use when the answer is not sitting ready on the screen.

Official sources

This article draws on the OECD's PISA 2025 Results, Volume I: Future-Ready Students, including the section on student life, digital tools and AI use, and the OECD press release published on 8 September 2026.