Hand, Head, Hours, Heart

Four short records that show a human was there when the machine did the work: what stayed human, what was learned from the miss, where the saved time went, and who carries the consequence. When somebody asks later whether this can be trusted, the answer is not a speech. It is four receipts.

Hand: what stayed human#

Write it down, even as a single line in a log. The hiring final call made by a named person after a second-round panel. The mask fit verified by two named engineers. The pricing decision signed off by a named director after the model disagreed.

Hand is not a claim to virtue. It is a record of custody. If something goes wrong, people know where to look for the person who will answer. If it goes right, people know where the courage sat. Both matter, and organisations tend to record neither.

Head: what you learned from the miss#

Every system will be wrong. What matters is whether the loop closes somewhere a human can see it. Capture one miss and one fix per cycle. The model over-counted a same-name subsidiary, so a registry check went into the pipeline. The prompt produced a tone that read as blame, so the instruction changed and the apology was rewritten. Triage mis-flagged repeat attenders, so a feature for prior admissions was added and tested against last quarter.

Head turns error into a changed policy instead of an anecdote. It is also the only one of the four that builds capability rather than recording it, and the reason it works is the reason set out at the after action review: reviewing the process, in writing, soon enough to be accurate.

Hours: where the time went#

If you cannot point to the investment, you did not make it. Sixty minutes released by an automated report, spent calling two clients back. Ninety minutes a week moved to rehearsing the handover that matters. Thirty minutes daily on an audit of yesterday's recommendations.

Hours proves that time is a resource being allocated rather than a puddle that throughput soaks up. It is the receipt that most organisations cannot produce, because the saved time was never reallocated to anything. It went back into volume, which is the mechanism described at usage theatre.

Heart: who bears it, including the unseen#

The first three receipts are all about the decision maker. What I kept, what I learned, where my time went. None of them records the people on the other end of the decision, and the ones furthest from the room are the ones least likely to appear in any system as data.

Heart records who carries the consequence. Named where they can be named: the customer pressured after a bereavement, the applicant whose file never reached a human, the worker whose shift pattern the optimiser rewrote. And explicitly including those who never enter the system to be counted at all, the people, animals and places downstream of a supply chain or a policy, whose harm arrives as somebody else's externality.

The test is simple and uncomfortable. Who is affected by this decision who has no way of appearing in the data it was made from? If the answer is nobody, check again. If the answer is somebody, the record says who, and whether anybody asked.

Provenance. Hand, Head and Hours are set out in SuperSkills (Kogan Page, 2026), chapter seven. Heart was added in September 2026, after a conversation on the Solutionary Voices podcast with Zoe Weil of the Institute for Humane Education, which raised the harms a decision maker will never know about. The formulation here is Rahim Hirji's and first appears on this page. The book carries three.

What it looks like in practice#

Clinical triage. A hospital introduces a tool that flags possible pneumothorax on chest X-rays. Hand: the consultant radiologist signs each critical finding and calls the emergency physician directly for the patients most at risk. Head: after a fortnight the team notices under-detection in a specific group, updates the training cohort and publishes the miss and the fix at governance. Hours: time released by a faster first pass funds same-day multidisciplinary discussion for borderline cases, so the dividend buys depth rather than volume. Heart: the patients in the under-detected group are named as the ones who carried the cost of the miss, and the review says who told them.

Customer operations. A bank predicts who is likely to leave and drafts retention messages. Hand: a manager keeps the right to cancel any message where a life event makes retention pressure unethical. Head: the team finds the model over-pressures customers after bereavement, changes the feature set and logs the change. Hours: time saved on drafting goes into phone calls to the highest-risk customers. Heart: the record names the bereaved customers who received the earlier messages, and what was done about it.

Product design. A team uses a model to propose an onboarding flow. Hand: the designer keeps the right to block any design that nudges users into consent they would not freely give. Head: the model underestimated drop-off for neurodivergent users, so a slower path was added. Hours: time saved on wireframes funds a full afternoon of moderated testing rather than an hour. Heart: the users who would have been nudged are named as the people the blocked design would have cost, and they are in the record even though the design never shipped.

Using it as an assessment#

These four are the closest thing in this body of work to an assessment of judgement, and the reason is that they record behaviour rather than asking a person to rate themselves. A person who has been exercising judgement can produce the receipts. A person who has been approving cannot, and the gap shows up without anybody having to make an accusation.

Two numbers fall out of the record over time and neither needs a new instrument. The proportion of decisions where Hand names somebody, and the trend in Head entries: a team that stops finding misses has usually stopped looking. The team-level version of the same practice, three questions at the end of any meeting, is at the pause minute.

5 to 8 October 2026: what a watermark cannot show#

Two reports in two days marked the limits of proving human involvement from the machine’s side. OpenAI said on 5 October that it will watermark eligible ChatGPT and Codex text in the European Union, and set out what the mark cannot do: “A watermark does not measure human contribution. It can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.” It added: “The absence of a detected watermark does not prove human authorship.” In the company’s own test on 400-token passages, replacing a quarter of the words with synonyms cut detection to 17 per cent, and the detector is not being made public at launch. Anuj Chopra of AFP reported on 6 October, in a story read at Philstar, that “A growing number of universities, from Yale to Cornell, have either banned or discouraged faculty reliance on AI detection tools as the primary evidence of alleged cheating”. He quoted the dean of Harvard College, David Deming: “We would all benefit from getting out of the AI-detection business.”

The record a reviewer sees can fail as well. METR reported on 6 October a flaw, found by a researcher in about ten minutes and patched within a day, that “could have allowed an agent to arbitrarily modify what the reviewer sees” in the viewer used to read transcripts of an agent’s work. The underlying record was untouched, and METR adds: “We have not observed agents exploiting this in any of our evaluations”. Its conclusion is that the systems that record and display an agent’s outputs “should be considered security-critical infrastructure”. A Hand line is only as good as the place it is kept.

If proof cannot be recovered from the output, it has to be made at the time, and that is the argument of this page. A position paper posted on 5 October by Yue Zhao and six colleagues puts the condition formally: “to be checked, a claim about an agent must first name its policy, its scope, the records that would settle it, and who writes them.” Their first example of such a claim is “a person approves every external email”. One survey suggests how rarely the record is shared. Laurel Kalser at HR Dive reported on 8 October that 78 per cent of more than 300 US managers surveyed by Highwire, a coaching programme, had used AI in the past year “to handle employee feedback or a performance review”, while 16 per cent of more than 700 individual contributors “said they were told AI played a role in their most recent review”. It is the programme’s own survey with no method published. None of this tests the four receipts, and the limit stated below stands.

9 to 15 October 2026: the proof moves from the output to the record#

Four reports in one week showed people trying to tell, from the output alone, whether a person did the work, and failing. Dan Milmo of the Guardian, in a report read at AOL on 14 October, covered the Institute of Student Employers’ annual survey: 92 per cent of employers suspected or confirmed misconduct during assessment and selection, up from 85 per cent the year before and 78 per cent the year before that, and 66 per cent reported at least one incident of a graduate misusing AI during an interview. A recruiter, Chris Morrow, said of such candidates: “It’s like they are reading off an Autocue.” Brandon Vigliarolo at The Register reported the same day a comedian who sent his own AI representative to an AI interviewer; it misstated his qualifications, offered to take around $60,000 against the $140,000 to $150,000 he wanted, and booked a follow-up interview with the other bot while he was out of the room. Nobody in that exchange had done any work, and nothing in the transcript showed it.

The universities gave up on reading it from the output. Georgia Luckhurst of Times Higher Education, in a report carried by Inside Higher Ed on 15 October, reported that Newcastle University will stop using Turnitin after the 2026 to 2027 academic year and the University of Westminster will retire it in July 2027, following Southampton. Westminster’s reason, in its words: the detectors are “probabilistic rather than conclusive and carry a recognized risk of false positives”. The evidence on detection has said so for two years; what changed is that institutions are now acting on it.

The same limit was reported for machines. SecurityHQ reported on 15 October, from 204 cybersecurity leaders, that only 49 per cent “could fully reconstruct every action an AI agent performed over the previous 30 days”, and that 99 per cent named at least one category where human review should remain mandatory, led by financial approvals at 61 per cent. Anthropic, in its report of 9 October on unintended actions by its own models, wrote that “a model’s own account of its reasoning is not necessarily reliable evidence of its beliefs or reasons for action”. If the output cannot show who did the work, and the agent cannot be trusted to say, the record has to be made at the time by the people who were there. That is the argument of this page, and the week supplied four more reasons for it. It did not supply a test of the four receipts, and the limit below stands.

What this has not been shown to do#

Nothing here has been tested. No study has compared teams that keep these records against teams that do not, and the framework has not been validated as a measure of judgement, capability or anything else. Records can be performed: a Hand line can name a person who rubber-stamped, and a Head entry can describe a miss that cost nothing. The practice makes judgement legible where it exists. It does not create it, and it will not by itself catch an organisation determined to look compliant.

Source#

Rahim Hirji, SuperSkills: The Seven Human Skills for the Age of AI, Kogan Page, 2026, chapter seven. Heart is an addition of September 2026 and is not in the book.

Related SuperSkills research#

• The three terrains
• The pause minute
• What is custody
• Augmented Mindset