The AI IQ trap: A high score won’t stop costly mistakes
An AI model drafts a confident-sounding reply in seconds. But is it right, does it fit this customer, and who takes responsibility? A new AI IQ comparison tries to put language model capabilities on a familiar human scale. The top score is about 136. It looks like a verdict. In practice, it is a prompt for a better question: what should people do as models become more capable?
A number that sounds like more
A score like 136 is easy to repeat. It brings to mind school reports, aptitude tests and league tables. That is exactly why it is worth taking a closer look before drawing conclusions about staffing, customer service or the future of your work.
The AI IQ website estimates scores for more than 50 language models and places them on a standardised bell curve. It combines 12 benchmarks, meaning standardised tasks used to compare performance, across four areas: abstract, mathematical, coding and academic reasoning. The overall score is the simple average of those four sub-scores. A model needs results in at least two areas to receive an estimated score at all.
As of mid-May 2026, GPT-5.5 leads the models covered, with an estimated score of around 136. GPT-5.4, Opus 4.7 and Gemini 3.1 Pro score roughly between 131 and 132 in the same comparison. These are results from a particular analysis, not a measurement of general human intelligence. So what does one number tell you when it combines four very different kinds of task?
Where the IQ comparison falls short
A model can get a tricky exam question right and stumble over an everyday task that looks simple. That unevenness is easy to lose when a range of results is boiled down to one overall score. The four areas say something about specific performance under test conditions. They do not automatically tell you how reliable the model will be on a phone call, in a warehouse or in a conversation with an unhappy customer.
Critics point to this gap between strong benchmark results and mistakes on specific tasks. VentureBeat cites the claim that GPT-5.5 has reached saturation on MMLU-Pro, while often failing on ClockBench. This should not be read as a general verdict on the model. It does show why a good average does not rule out weaknesses.
The method also deserves scrutiny. AI IQ lists the 12 benchmarks and shows calibration curves, but according to VentureBeat it does not publish the raw data or the precise mathematical conversions as open datasets. For you, the score can start a conversation. It cannot replace testing with the tasks your business actually handles. That is where the gap between impression and performance can prove costly.
Progress does not wait for your annual review
You may be waiting for a quieter quarter to look at AI. The trouble is, progress does not follow your project calendar. The AI IQ time series records an increase of about 60 points in estimated IQ for top models between October 2023 and mid-2026. That is the pattern in this particular series, not a promise that the next jump will be just as large.
The technical foundations are also growing quickly. Epoch AI estimates that the amount of AI computing capacity sold, measured in H100 equivalents, is increasing by a factor of 3.3 each year. That amounts to a doubling roughly every seven months. The five largest chip designers had, according to the estimate, shipped around 20 million AI chips in total by the fourth quarter of 2025. These figures help explain why more computing power is becoming available for models. On their own, they do not predict which capability will improve next.
Let us run a thought experiment. Suppose leading models reached an IQ-like score of 180 to 200 by December and around 250 in the quarter after. This is not a forecast, it is an exercise: if AI becomes far faster and better than almost anyone at many thinking tasks, our value shifts. Away from knowing and calculating, towards good questions, judgment, responsibility and trust.
The figures remain a rough translation of human IQ scales to model tests. Even so, the practical question for you is the same: which decisions can AI help prepare faster, and which need to stay with people?
What AI cannot decide for you
When a machine produces a better draft, responsibility for using it still rests with your organisation. A retailer might use AI to check refund requests. But whether a long-standing customer deserves a goodwill gesture after a mistake may depend on a relationship the data cannot see. A law firm can sort documents by priority. That still leaves the decision about which uncertainty needs explaining to the client.
Taste matters too. Which wording sounds like your business? Which offer feels fair? When should a team member escalate an unusual situation rather than follow the standard process? These are not minor extras. They determine whether a quick suggestion holds up in real work.
McKinsey estimates that current technologies could automate tasks accounting for more than half of working hours in the US. That does not mean a corresponding number of jobs will disappear, McKinsey says. More than 70 per cent of the skills employers seek are relevant to both automatable and non-automatable work; a further 12 per cent are currently needed only for work that cannot be automated. What your business needs will therefore also depend on how well people work with the new tools.
What this means for your business
Do not start by asking whether AI is cleverer than a person. Pick a specific task from last week. Perhaps it was the email that three colleagues had to agree on before it went out. Or the appointment request that only got a reply after hours. Decide whether a model could draft a response, sort information or suggest a next step.
Next, set a clear point for review. Who checks the content? Which mistakes are acceptable, and which are not? A slightly off note in an internal summary may be irritating. A promise in a contract, a hiring decision or medical advice calls for a different standard. Also write down when the task must go to the person responsible.
A practical way to start over the next few weeks:
• Choose a recurring task with manageable consequences, not the most sensitive process in the business. • Ask staff to test the model on typical cases and difficult exceptions. • Check results against your own quality standards and record mistakes. • Name someone to update the rules and checks when the workflow changes.
In a one-person business, you may be the one doing the checking. In a small or medium-sized business, handovers between the business team, IT and management need to be clear. A larger organisation also needs shared rules, so each department does not draw its own line between a suggestion and a decision. The next sensible step is not a major project. It is one specific task with one person accountable. But who needs new skills first?
People need to learn faster, not know everything
The new work often starts with a modest skill: describing a task clearly enough to get a useful answer from a model. Type only “Write a customer email” and you may get a polished message that misses the crucial context. Include the goal, background and constraints, and you have better suggestions to check. That calls for subject knowledge and a willingness to refine the question.
McKinsey identifies problem framing, checking AI outputs, interpreting results, handling exceptions and deciding when to escalate a case as skills for working with AI. It also says job postings mentioning AI skills rose nearly sevenfold in two years. For a small team, this does not have to mean a comprehensive training programme. Reviewing real examples together can show where people are trusting too much or not trying enough.
Picture next Monday: a colleague checks an AI draft for missing context, not just spelling. A team member flags an exception instead of quietly waving it through. And everyone knows who makes the final decision. This way of working does not depend on an IQ score. It will still serve you when the next score on the chart rises.
Frequently asked questions
What does an AI IQ score of 136 mean?
The score is an estimate from a particular comparison of language models. AI IQ combines 12 benchmarks across four areas and calculates an average. It places test results on a human IQ scale, but it does not measure general human intelligence or directly predict how reliably a model will handle your specific business tasks.
Can AI really have a higher IQ than a person?
An AI IQ score is an analogy based on standardised tasks, not a direct measurement like a human IQ test. Models can perform very well on some tests and make mistakes on others. A higher estimated score does not automatically mean a model is better than people at judgement, responsibility or everyday experience.
How quickly are AI models getting more capable?
According to the AI IQ time series, the estimated score for top models rose by about 60 points between October 2023 and mid-2026. That describes this particular series and is not a forecast. More computing capacity also does not guarantee that every capability will improve at the same rate.
Which tasks should businesses delegate to AI?
Start with a recurring task where the consequences of a mistake are manageable. A model might sort information or prepare a draft. Decide who checks the result and which cases must be passed on. Decisions with legal, financial or personal consequences need clear accountability and appropriate checks.
What skills do employees need to work with AI?
Useful skills include framing a problem clearly, checking results, judging uncertainty and handling exceptions. Employees also need to recognise when a decision belongs with a responsible person. These skills are not built through theory alone. Practising with real tasks from the business helps reveal where people need support.
Interested?
Want to know what this means for your company?
Sources
- AI IQ is here: a new site scores frontier AI models on the human IQ scale. The results are already dividing tech. · VentureBeat, 2026-05-13
- Global AI computing capacity is doubling every 7 months · epoch.ai, 2026-01-09
- What you need to know about AI chips · epoch.ai, 2026-07-14
- Human skills will matter more than ever in the age of AI · McKinsey & Company, 2026-01-06
This article was researched and written with AI assistance and automatically checked against the sources listed before publication. More on our AI transparency
Klarframe AI Readiness Check
How ready is your company for AI?
About ten minutes, instant result with your stage, a 90-day plan, suitable solutions and notes on the EU AI Act.
