AI did not remove the complexity of software. It removed the interface that told you where the complexity went, and who is now paying for it.
Larry Tesler is the reason you can cut, copy, and paste. At Xerox PARC in the 1970s he and Tim Mott built Gypsy, the first editor where text stayed editable without entering a command mode first.
Steve Jobs recruited him to Apple in 1980, where he stayed 17 years and rose to chief scientist, driving a car whose license plate read NO MODES. He spent that career deleting the states that confused people, a UX career before the field had a name.
By the mid-1980s he had turned that instinct into the law of conservation of complexity, which says the hard part of any application cannot be removed, only handed to somebody (or something) else.
“Every application has an inherent amount of irreducible complexity. The only question is who will have to deal with it — the user, the developer, or the system.” — Larry Tesler
Forty years later the chat box arrived and looked like an answer: type what you want, the machine does the rest, no menus, no modes, no manual.
It was not an answer, it was a transfer, in three directions at once.
Some moved into the model, which is real progress. Some moved onto the user, who now writes the specification and judges the result. The rest went to your team, into evaluation work nobody staffed.
Tesler’s law did not get repealed. Moving difficulty is the job of design, so the transfer is not the failure. Concealing it is.
Eight points on where it went.
The Law Nobody Repealed
He names three parties where most UX writing collapses the question to user versus system. Here’s an example for today’s world:
The user
- Then. Learned the menus, the modes, a shortcut or two.
- Now. Writes the instructions and then reviews whatever comes back.
The application developer
- Then. Conditional logic, 100 error strings, every edge case anyone caught.
- Now. Evaluations, retrieval plumbing, and prompts that quietly became product logic.
The platform developer
- Then. Microsoft or Apple, on a release cycle you could plan against.
- Now. Your model provider, changing what it absorbs every few weeks without telling you.
This column is new, because a role appeared that Tesler had no reason to name.
The reviewer
- Then. Nobody, since the person doing the work was the person checking it.
- Now. Certifies output nobody in the room produced.
Four columns means a ledger, and this is how design has always worked. Every clean screen you admire was paid for out of one of those columns, and the screen never says which. Simplicity in an interface is almost always someone else’s workload.
Simplicity in an interface is almost always someone else’s workload.
The difference now is speed. The transfer moves more at once, and leaves no mark on the interface where it landed.
Count the Receiving Side
A ledger is good news, because it makes the whole thing countable. If a feature saved the user 10 minutes, those minutes showed up somewhere you can already instrument: review time per artifact, evaluation hours your team now spends, inference cost per task. Almost nobody counts them, which is why the savings look free.
Almost nobody counts them, which is why the savings look free.
That is a gap in practice rather than a limit of the technology. It closes the first quarter somebody measures both ends of a transfer.
Which leaves an answerable question nobody has asked out loud. Menus and modes were bounded, learnable, and the same on Tuesday as on Monday. Writing a specification and carrying the decision are neither.
We may have spent 40 years taking work off the user and three handing some of it back.
The Text Box Is a Transfer
Jakob Nielsen called generative AI the first new interaction model in 60 years, in AI: First New UI Paradigm in 60 Years. Users state the result they want instead of the steps. He is right about the shift, but telling a system what you want requires knowing what you want, in detail, in words. That is a specification, and writing one used to be a professional skill with training and review.
It is now the control surface for hundreds of millions who never asked for the job.
Amelia Wattenberger named the cost early in Why Chatbots Are Not the Future. A text field looks identical to a search box, a login form, and a credit card field. Users can learn which prompts work, she wrote, but the burden sits with every user. A blank text field is the most flexible control ever shipped and the least instructive.
A blank text field is the most flexible control ever shipped and the least instructive.
A dropdown teaches you the domain while it constrains you. A disabled button shows a rule you would otherwise discover by failing.
Verification Is the Bill
In July 2025, METR ran a randomized controlled trial with 16 experienced open-source developers across 246 tasks in repositories they had maintained for years. The result, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, was a 19 percent slowdown when AI was allowed. They forecast a 24 percent speedup going in and estimated 20 percent coming out.
Sit with that gap. The finding is not that the tools failed, it is that professionals could not feel where their own time went.
METR has since labeled those numbers historical, and redesigned a 2026 follow-up after too many developers declined to work without AI. Whatever the speedup is now, the tools are foundational.
Generating feels like progress in a way that verifying never will.
The work did not disappear. It changed from producing to checking, and checking is harder to see.
So design for the checking. Show a diff instead of a wall of replacement prose, and put the source where the decision gets made.
Where AI Absorbs Real Complexity
Here is the part the critique skips. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied 5,172 customer support agents for Generative AI at Work in the Quarterly Journal of Economics, 2025. AI assistance raised issues resolved per hour by 15 percent, and less experienced workers improved in both speed and quality while the most experienced saw small declines.
That distribution is the finding.
The system absorbed complexity that took years of tacit knowledge: the phrasing that calms an angry customer, the judgment about which policy applies. For an expert who already carried it, nothing was left to take.
Geoffrey Litt made the wider argument in Malleable software in the age of LLMs. The barrier to end-user programming was never desire, it was turning a rough idea into code.
AI is the first tool in decades that raised the ceiling on how much complexity the system side can hold.
That is a change to the equation Tesler wrote and it is also the outcome he spent a career asking for.
The Deterministic Floor
Some complexity does not move at all.
Payroll is the clean case, where a paycheck is either correct or it is a legal problem, because withholding is a rule with a citation and an audit trail. Ninety-nine percent correct payroll is failed payroll. The same holds for authorization, tax filing, reconciliation, dosing, and safety interlocks.
Stanford RegLab tested the best case in Hallucination-Free?, published in the Journal of Empirical Legal Studies in 2025 on products as they stood in May 2024. The tools were retrieval-grounded and built for lawyers by LexisNexis and Thomson Reuters. They hallucinated on 17 to 33 percent of queries.
The floor has a second half, and enterprise software sits on it. An underwriter, a claims adjuster, a clinician: their work is hard because the domain is hard, and no interface reduces that.
AI can retrieve the rule and draft the letter, but it cannot own the decision, because owning it means being the person asked why a year later. That accountability is a ledger entry, not a rounding error. The model can hold the search, never the settlement.
The model can hold the search, never the settlement.
Let the model prepare the change, then let deterministic code execute it and a named person approve it. That boundary is knowable before a line of code exists.
The Jagged Edge
Above the floor the boundary is not knowable in advance, absorption is uneven, and the unevenness stays hidden. Stanford HAI’s 2026 AI Index puts generative AI in at least one business function at 70 percent of organizations, while agent deployment stays in the single digits.
Tesler named the reason decades earlier. In Tesler’s Theorem and other adages and coinages he gives his version of what others call the AI Effect: intelligence, he wrote, is “whatever machines haven’t done yet.” Whatever a machine absorbs stops counting, so the edge moves without anyone announcing it.
I catalogued the same edge by artifact in A UX Designer’s Field Guide to What Survives. Ask for a research readout before the research exists and you get 10 pages in 20 seconds, every quote invented.
The user’s new job is knowing where the capability ends, and nothing in the interface tells them.
The old interface at least failed out loud. A model answers everything in the same steady voice whether it knows or not, a defect rather than a limitation, because presentation closes it.
Mark Steyvers and colleagues measured the calibration gap in What large language models know and what people think they know, Nature Machine Intelligence, 2025. Longer explanations raised confidence without improving the answer; aligning them to internal confidence narrowed it.
Allocation Is a Design Decision
Eric Horvitz wrote the decision procedure 27 years ago. Principles of Mixed-Initiative User Interfaces, presented at CHI 1999, treats each automated action as a judgment call with three inputs:
- How uncertain the system is about the goal
- What a wrong guess costs
- Whether the person can cancel directly
The second input sets the bar. When the price of being wrong is a corrected paycheck, the answer is never a generated one.
Saleema Amershi and 13 colleagues published Guidelines for Human-AI Interaction at CHI 2019, validated with 49 practitioners, and Horvitz is on that paper too. The first two guidelines are the ones teams skip: say what the system can do, and how well it does it. You are allocating complexity whether or not you write it down.
You are allocating complexity whether or not you write it down.
The default allocation is a blank box, a confident answer, no capability signal, no cheap way to verify. It is the cheapest option for the team building it, which is the trade Tesler spent his career arguing against.
Conclusion
Tesler was arguing about menus and modes, and the stakes were minutes. The stakes now are whether a person can tell a correct answer from a confident one, a heavier thing to hand someone through a rectangle.
His law holds up anyway. What changed is the size of the container on the system side, and that change is real. Models absorb translation, tacit knowledge, and boilerplate no interface could carry before, right up to the deterministic floor where they stop. Tesler would have taken that trade in a second.
What also changed is that the transfer stopped announcing itself. A cluttered toolbar declared its complexity, loudly and badly, but it told the truth about the cost. A blank prompt says nothing, and the burden arrives unposted.
Your job is the ledger. Decide, in writing, which of the four columns each piece of difficulty belongs in, then build the controls that hold that decision in place. Constraints that teach. Capability signals that admit where the ground drops away.
The complexity is still there, and it is still conserved. The only thing you choose is who pays for it, and you are making that choice every single sprint, named or not.
Action Items
None require a model change.
- Post the ledger entry first. Name the column that absorbs the irreducible complexity, where reviewers will see it.
- Instrument the receiving column. An unmeasured transfer is indistinguishable from an improvement.
- Turn one text field into a control. Where users type the same instruction repeatedly, give it a selector. A constraint teaches while it limits.
- Budget verification time, not generation time. Count the seconds spent checking, not just the seconds saved.
- Design for the novice gain. The measured wins concentrate among less experienced users.
- Draw both halves of the floor. Mark which outputs must be exact and which decisions need an owner. The first gets deterministic code, the second a named approver.
- Give the jagged edge three treatments. A source chip for retrieval, a visible hedge for inference, a refusal out of domain.
- Audit against guidelines one and two. Check whether a shipped feature states its limits anywhere a user would look. Most do not.
Tesler’s Law: complexity moved to a different part of the journey was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.
