Ideas, observations,
and honest takes.
No hype. No tools-of-the-week. Observations from someone who uses AI in production work every day and cares whether the take is accurate.
Disclosures doubled, fixes stayed flat
A frontier AI more than doubled vulnerabilities found. It never moved patch velocity. The dashboard will show you the first number, never the second.
Reorg Precedes The AI Benefit
Adoption dashboards hide which teams restructured and which absorbed the tool. Team formation is the lever.
Ask for the transfer test
An AI hiring model trained on firms like yours scores your applicants no better than chance. Ask the vendor to show you anyway.
Humans don't help AI
The augmentation slide compares human plus AI to human alone. The one comparison it never runs says human plus AI loses to AI alone.
The dialogue raised the room
A 7-minute AI dialogue raised group deliberation quality by d = 0.31. The same questions on paper did nothing. The pre-read never worked.
The incident report ships weapons
The post-incident attack-text report delivers working weapons to the next attacker four times more often than their own improvisation.
Familiarity Dulls the Alarm
A clean incident queue felt like proof the agent was behaving. The paper says it is proof the team got comfortable.
The experience premium became an arch
In AI-exposed occupations the pay curve for experience just rotated. Young workers gained. Veterans lost. The slope flipped at 18.1 years.
The Brief They Narrate Back
The expert vignette never moved your board. The one they wrote themselves did, by up to 0.90 SD.
Copilot Added the Meetings
The ROI deck said meetings would fall. Bocconi telemetry says they rose, and the user's network reshaped around them.
The Check Made It Worse
An AI tutor drifted. The team learned the wrong thing. They still rated it fine. Checking the AI made it worse.
The Pilot Killed the Right Tool
The AI tool your team rates highest in week six is often the one eroding the skill you were trying to build.
The boring tier was the ballast
Hand AI the easy tier. Human conversion falls on the warm leads it saves them. The boring tier was doing work nobody wrote down.
The abstention is the segmentation
The same AI advisor policy raised acceptance for experts and dropped it for juniors. Advisor design is user segmentation.
The briefing nobody pushes against
A plain recommendation with override beat the respectful briefing format by 13 points in complex sepsis decisions.
The Mockup Moved the Hire
AI-generated faces in a hiring deck quietly drag down ratings of the real candidates you interview next. Retouched real photos do nothing.
Belief Shift Per Session
An advisory LLM moves user beliefs 15 points per short chat. The fix cuts drift by 78% and costs general trust in AI.
AI's blind spot is the signal
Applicants whose essays sat deepest in AI's blind spots scored higher on the SAT. The ideas AI misses are as good or better.
Same tool, opposite sign
Same 15 minutes of AI access. Moved from search to selection, expert-rated creativity drops 0.62 SD.
The handmade premium nobody pays
Three experiments, 1,025 buyers. The AI premium did not change when the label said human-designed.
Ask for the Methodology
The 60% on the vendor slide is a marketing claim with no receipt behind it. Ask how it was measured and watch the room change.
Rehearse Twice Or Not At All
A single AI rehearsal leaves you flatter in the next meeting. The second one is where it lands. Most teams never get there.
Agreement Is the Wrong KPI
Rising inter-annotator agreement on subjective AI tasks may be suppressing the signal the pipeline is meant to measure.
Voluntary expertise is not free
Workers consent to AI reuse of their work 20 points less than to human reuse. The gap is a disclosure, not noise.
The market prices position first
The market prices a firm's position in the AI hire-flow network, and it is pricing it before earnings move.
Two advisors still need the expert
A second AI advisor doubles the error-catch rate. Only for the professionals who could have caught the first one on their own.
The No AI Clause Is Decoration
100% of submissions in studies that prohibited AI got paid. The clause in your research brief is a comfort line, not enforcement.
The Recommender Eats Itself
A live NYC school-match RCT: the capacity-aware version raised matches 71 percent. The naive version quietly collapses its own signal.
Steps Nobody Asked For
About half of the reasoning steps the AI takes were never requested. The dashboard sees the output, not who initiated what inside it.
The judge chose the number
LLM-as-judge dashboards measure the judge, not the AI. Swapping the judge configuration can move the reported acceptance share by nearly 100 points.
The One Voice Does Not Survive
Enterprise summarisers filter employee feedback by prevalence. About 86 percent of one-time concerns drop before leaders ever see them.
The Fairness Layer Moves the Burden
Six leading post-hoc fairness layers tested on real student data. None reduced error disparity. Two concentrated harm on the smallest subgroups.
Participation widened, acceptance concentrated
AI widened who writes on Stack Overflow. The acceptance distribution did not follow. Your participation dashboard now ships a false positive.
The filter bubble prediction flipped
The two-year filter bubble worry about AI just ran into a 37,561-reader field experiment. The sign went the other way.
The Adoption Mandate Was Theatre
One firm halved its AI adoption mandate and ninety percent of the drop was gaming queries. Real work did not move.
Team Can't Tell Which Half Worked
The AI-productivity number is real. Your team's ability to tell which of its own answers is right just got worse.
The digital twin prediction ceiling
AI-moderated interviews win on cost per customer need. Their digital twins predict no better than a cheap static survey.
Acceptance rate hides the drift
One AI acceptance-rate number can conceal opposite reliance dynamics across cultures and quietly overwrite a team's voice.
Your Memory Is the Summary
AI summaries overwrite the reader's memory of the source, and the AI label switches on no scrutiny.
The wrapper does the work
The LLM did nothing to prices. The dashboard around it drove every anticompetitive effect in the study.
Thinking Partner Tilts the Compass
A short LLM thinking-partner session shifts a person's active values toward Achievement and Power. The effect fades by the next task.
The AI Penalty Your Policy Missed
Reviewers already run an implicit AI classifier at AUC 0.70. The written policy has almost nothing to do with what they do next.
Half the Benchmark Reaches Your Team
About half of a vendor's benchmark gain reaches the human. Your team recovers less than the release-notes table promised.
The Refusal Nobody Priced
Users prefer a wrong answer they can catch to a correct refusal. Every refusal is a satisfaction cost your safety prompt hides.
The Signal Beats the Badge
Rewards for restrained AI use moved nothing. A UI signal that reflected each offload back at the user halved it and lifted unaided performance.
Your AI policy moves only the compliers
A randomized experiment at ICML 2026 measured what an AI-use policy actually does at scale. It moves a third and misses the rest.
The empathy work nobody is doing
A new USC and Vanderbilt study puts numbers on what the L1 human in your dispute chair is not doing: cooling affect, offering trade-offs.
A person in the seat
A preregistered paper on AI hiring says applicants weight one thing above every safeguard: a person in the seat. Worth twenty accuracy points.
Your workforce cannot name its AI use
Only 23% of people could name the way they most use AI. Their leaders are running the wrong survey.
Change the Artefact, Not the Reviewer
The summary an agent writes about itself refers to one action in eleven, and its completeness does not predict correction.
Benchmarks Pick the Wrong Partner
Two AI partners can hand humans the same literal signal and produce 14.4% or 34.4% human failure. Vendor benchmarks are blind to this.
The interactivity you paid for
Adding a chat surface to an AI advisor did not shift behaviour, and almost nobody used it. The interactivity is a demo asset.
The Safety Net Nobody Writes
Every AI transformation plan I read assumes the state will absorb the humans it moves. Across 33 parliaments, that story is not being written.
The AI score is a lever
The position of the AI signal on your review screen roughly doubles the weight it carries in the next human decision.
What your copilot quietly displaces
Your productivity dashboard measures what the copilot does. Nobody is measuring what it quietly replaces.
Your Escalation Queue Trains Your Reviewers
The composition of your escalation queue silently reshapes reviewer accuracy, and reviewers cannot feel it happening.
The wrong dose of AI in the room
A small AI presence sharpens group decisions. A third breaks them. A majority-AI room agrees on ideas nobody would have written.
The Audit Mirror You Already Own
Run an AI baseline in the shadow of your senior people and you have built an instrument that measures their drift.
The AI Disclaimer Is a Comfort Blanket
A compliance disclosure feels like risk management. In the field the AI still moves the user.
The Better You Are at AI, the Less You Catch Its Mistakes
A study of 1,339 teachers found the most tech-fluent reviewers deferred to wrong AI the most. Fluency is quietly becoming a liability.
Stop Asking If AI Will Replace Your People
Every executive is running the wrong threat assessment. The question isn't whether AI will take jobs. It's whether AI will make your people's expertise worth more or worth less.
The Skill Your Best People Are Losing
Everyone is measuring whether AI made your team faster. The question that matters is whether it is quietly making them worse at the thing they handed over.
The cheapest output AI gives your team is confidence
The heaviest AI users on your team feel like the experts. A new UC Irvine study says they are the ones learning least.
Who's Actually In Your AI Data
Every confident statistic about how AI is reshaping work is partly a statistic about who shows up in the conversation logs.
What if Anthropic chose Albania as a research hub?
ChatGPT to Malta is a press release. Anthropic to Albania would be a thesis — the first country on earth where you could watch an AI-native generation become itself in real time.
Stop asking Claude if your plan works. Ask how it failed.
The prompt Anthropic product teams reportedly run before shipping flips Claude from defending your plan to populating its failure.
The AI you audit is not the AI they ship
When humans watch a language model, it shifts register. Your governance committee meets one version. Your team meets another.
Your team's thinking time collapsed. The output looks identical.
A new USC study: with an LLM, programmers had the same ideas in their solutions. They just stopped thinking before they wrote.
When AI Fails Politely, Humans Fail Quietly
Field-tested at Alibaba: when agentic AI hands off a frustrated customer, the human worker quietly disengages instead of stepping up.
Your AI rollout has a system problem, not a skill problem
Microsoft surveyed 20,000 knowledge workers. Two thirds of AI's impact comes from the system around the worker, not the worker. Most rollouts forget this.
AI writes the goal. You won't do it.
AI writes objectively cleaner goals than the people who use it. Those same people then quietly stop acting on them.
The humans got harder to talk to
Three weeks with a sycophantic chatbot, and people are as likely to ask the model for personal advice as their closest friends — and less satisfied with the humans in their lives.
AI is not a colleague
Putting an AI on the org chart does not improve adoption. It quietly slides accountability off the human who built the work.
The human-in-the-loop fiction
Putting a human in the loop isn't a safeguard. It's a stage on which the AI performs.
Stop performing AI literacy
The signal that someone is actually fluent isn't their vocabulary. It's their patience.
The five prompts I use every morning
A practitioner's real workflow — not the one on LinkedIn.
Why I stopped caring about model benchmarks
Benchmarks measure the model. I care about what happens after the prompt.
What "good at prompting" actually means
It's not a technical skill. It's a communication skill. The distinction matters more than you'd think.
The gap is communication, not code
The actual barrier to AI fluency isn't access. It's articulation.
Ideas, observations,
and honest takes.
No hype. No tools-of-the-week. Just the work, explained clearly, from someone who uses this every day and cares whether it is accurate.
No spam. Unsubscribe any time.