Taste Is the Scarce Thing Now

Twenty dollars per million tokens in November 2022. Seven cents per million tokens by October 2024. Stanford's AI Index puts the fall at "a more than 280-fold reduction in approximately 18 months", measured against a fixed capability bar: the cost of querying a model scoring the equivalent of GPT-3.5 on the MMLU benchmark. The specific numbers matter less than the shape. Something that used to be a line in a budget became a rounding error inside two years.

In our studio the practical effect arrived as a wall covered in options. Not a better option. Sixty of them, produced in an afternoon, of which four were worth discussing and none was finished. We had removed a constraint we had been unconsciously relying on, which was that making a thing took long enough that you only made the things you meant.

The Simon line, and the half that gets cut

The obligatory citation here is Herbert Simon, and it is almost always truncated. The source is "Designing Organizations for an Information-Rich World", in Martin Greenberger's edited volume Computers, Communications, and the Public Interest, Johns Hopkins Press, 1971, pages 37 to 52. The passage runs:

What information consumes is rather obvious: it consumes the attention of its recipients. Hence, a wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information resources that might consume it.

It begins earlier still: "In an information-rich world, the wealth of information means a dearth of something else: a scarcity of whatever it is that information consumes."

The half that gets cut is the clause after the famous one. Simon is not lamenting overload. He is stating a design problem: attention is now the scarce input, therefore the allocation of attention is the thing that has to be engineered. The essay's title is about designing organisations, not about how distracted we all are. Read whole, it is closer to an operations memo than to a cultural complaint, and it says that when a resource becomes scarce you build machinery for rationing it.

I have used the short version lazily for years. The long version is the one that tells a studio what to do.

What actually got cheap

Two things fell at once, and they are worth separating.

The cost of a unit of output fell, which is the AI Index number above. And the number of units produced rose, which is harder to measure. The most-cited attempt is by Alina Valyaeva at Everypixel Journal, published 15 August 2023, which estimated "more than 15 billion images created using text-to-image algorithms since 2022 to 2023", with people "creating an average of 34 million images per day", broken down as roughly 12.59 billion from Stable Diffusion across all channels, 964 million from Midjourney, 916 million from DALL-E 2 and around 1 billion from Adobe Firefly in its first three months.

Treat that figure with care. It is an independent estimate built by extrapolating from job-processing rates and repository download statistics, not an audited industry number, and the outlets that repeat it rarely say so. I quote it because it is the best public attempt I could find and because even if it is wrong by an order of magnitude the conclusion survives: the supply of plausible images is now effectively unbounded.

What did not get cheap is the thing that turns a plausible image into a decision: looking at it, comparing it against an intention, and being willing to throw it away.

The adoption data, and what it actually shows

Three surveys, none of them perfect, all of them public.

The UX Tools figure I find most interesting is the one nobody quotes. Its data was collected from 2,220 respondents between November 2024 and January 2025, and the survey reports "75.2% Percentage of AI usage focused on text-based tasks (text/copy, documentation, content generation) rather than visual design". Agency design leaders reported the highest adoption at 88.7%, corporate leaders the lowest at 78.2%. So designers are heavy adopters, but overwhelmingly for writing, documentation and research synthesis, not for making the pictures. The visual-generation revolution is, in measured design practice, still mostly a writing revolution.

The best single piece of evidence for the thesis

It is not an adoption number. It is a complaint. In the 2025 Stack Overflow survey, the leading frustration, reported by 66% of respondents, is "AI solutions that are almost right, but not quite". The follow-on, at 45.2%, is that debugging AI-generated code takes longer. Positive sentiment toward AI tools fell to 60%, from over 70% in the two preceding years, even as adoption rose.

That combination is the whole argument in one data point. Usage up, trust down, and the dominant pain is not that the output is bad but that it is nearly good. Nearly good is the most expensive kind of output there is, because it defeats cheap rejection. Obvious rubbish costs a second to discard. Something 90% right has to be read in full, understood, and then either repaired or refused, and the refusing is the part people find hardest.

Nearly good is the most expensive kind of output there is, because it defeats cheap rejection.

Production has been automated. Evaluation has not, and the volume of things requiring evaluation went up by whatever multiple you believe.

What a studio actually charges for

This is where the argument stops being atmospheric. Look at what the standard contract in the field is built out of. The AIGA Standard Form of Agreement for Design Services is modular, and of its five modules two are marked required: Basic Terms and Conditions, and Intellectual Property Provisions. The optional ones are the medium-specific sets, for print, interactive and environmental work. AIGA is explicit that this is "not a one-size-fits-all approach" and that "design firms develop their own custom proposal document for each project".

The medium is optional. The rights are not. What a studio sells, structurally, is a transfer of rights in a specified selection of work, plus the judgement that produced the selection. Not hours, and not files.

The legal position under generative tools makes this sharper rather than blurrier. The United States Copyright Office published Copyright and Artificial Intelligence Part 2: Copyrightability in January 2025. Its conclusions are narrow and useful. On prompting: "prompts alone do not provide sufficient human control to make users of an AI system the authors of the output", and "Based on the functioning of current generally available technology, prompts do not alone provide sufficient control." On what remains protectable:

Human authors are entitled to copyright in their works of authorship that are perceptible in AI-generated outputs, as well as the creative selection, coordination, or arrangement of material in the outputs, or creative modifications of the outputs.

The Office also concluded that "questions of copyrightability and AI can be resolved pursuant to existing law, without the need for legislative change".

Read those two findings next to each other. The generated material is not yours. The creative selection, coordination and arrangement of it is. Which means the only part of the work that is legally yours to license is precisely the part that consists of judgement. That is not a metaphor about taste being valuable. It is the thing you are actually permitted to sell.

How I would reorganise a small studio

Specifically, and with the admission that we have done about half of this and are bad at the rest.

Put a ceiling on options, not a floor

The instinct when generation is cheap is to produce more and pick the best. This inverts. Decide the number of options before generating, make exactly that many, and force a choice. An option budget of five with a rule that all five get argued about beats sixty with a rule that we will know it when we see it. Simon's point about allocating attention efficiently applies at the scale of an afternoon.

Name the person who decides, per project, in writing

Selection costs are hidden when everyone selects a bit. A studio of four people looking at sixty options spends more judgement than one person looking at five, and produces a worse decision, because the group converges on whatever is least objectionable. This is the single change that has helped us most and it costs nothing.

Keep a rejection log

Write down what was refused and the reason, in a sentence, at the time. This is the only artefact that accumulates into a house standard. Portfolios record what you accepted, which is much less informative about a studio's actual taste than what it refused and why.

Price the selection, not the production

If the contract's required modules are terms and rights, then the proposal should be organised around scope of use, exclusivity and duration, not around deliverable counts. Deliverable counts are now a bad proxy for effort, and clients have noticed. Quoting forty variations as though it were forty times the work of one is a position that will not survive contact with a client who has used the same tools.

Write down the standard before you generate against it

Evaluation is cheap when the criteria exist beforehand and ruinous when they do not. A one-page written brief that states what would make an option wrong, not just what would make it right, converts an afternoon of taste into an afternoon of checking.

Studio operating changes, in order of payoff

1. Option budget fixed before generation begins.
2. One named decider per decision, written in the brief.
3. Rejection reasons logged at the moment of rejection.
4. Negative criteria in the brief: what would make this wrong.
5. Pricing by rights and scope, not deliverable count.
6. A standing rule that "almost right" is refused, not
   repaired, unless repairing it is under a stated time box.

Rule 6 is the expensive one and the one we break

Studio operating changes, in order of payoff

1. Option budget fixed before generation begins.
2. One named decider per decision, written in the brief.
3. Rejection reasons logged at the moment of rejection.
4. Negative criteria in the brief: what would make this wrong.
5. Pricing by rights and scope, not deliverable count.
6. A standing rule that "almost right" is refused, not
   repaired, unless repairing it is under a stated time box.

Rule 6 is the expensive one and the one we break

Studio operating changes, in order of payoff

1. Option budget fixed before generation begins.
2. One named decider per decision, written in the brief.
3. Rejection reasons logged at the moment of rejection.
4. Negative criteria in the brief: what would make this wrong.
5. Pricing by rights and scope, not deliverable count.
6. A standing rule that "almost right" is refused, not
   repaired, unless repairing it is under a stated time box.

Rule 6 is the expensive one and the one we break

Four ways this thesis could be wrong

I would rather set these out than have them arrive as comments.

Taste might be automatable too

The argument assumes selection is harder to automate than production. There is no law requiring that. Ranking and filtering are, mechanically, more tractable problems than generation, and a system that reliably picks the best four out of sixty would remove the scarcity the whole thesis rests on. What it could not do, on current law, is own the result. But a studio betting its business model on a durable gap between generating and judging is betting on a gap nobody has proved is durable.

The adoption numbers are self-selected

Every survey quoted above is voluntary and skewed toward people already engaged with the topic. The Stack Overflow survey covers developers, not designers, and I have extended its findings to design work by analogy rather than by evidence. The UX Tools sample is 2,220 people who read a design-tools newsletter. None of this is a census.

"Almost right" may be a transitional artefact

The 66% frustration figure is a snapshot from one year. If output quality keeps improving, the nearly-good band narrows and the evaluation burden falls with it. A thesis whose best evidence is a current failure mode is vulnerable to that failure mode being fixed.

Clients may simply not pay for judgement

This is the objection I take most seriously. "Taste is scarce" does not entail "taste is purchasable". A market can value a scarce thing at zero if buyers cannot tell it apart from the abundant thing. The Copyright Office finding helps here, because it gives selection a legal existence that is independent of whether a buyer can perceive it, but legal existence and commercial demand are different things, and I have no data on the second.

What I would still bet on

Simon's sentence holds because it was a claim about allocation, not about volume. When something becomes abundant, whatever it consumes becomes the constraint. Images now consume attention and judgement, and neither of those got 280 times cheaper.

The version of this I am most confident in is narrow. It is not that taste is the new competitive advantage, which is the sort of thing that gets said in conference talks and means nothing operationally. It is that the cost structure of a small studio has moved, that the expensive hours are now the ones spent deciding rather than the ones spent making, and that almost every studio process in existence was designed when the opposite was true. Fixing the process is unglamorous and available immediately. The philosophy can wait.

Sources

• Herbert A. Simon, "Designing Organizations for an Information-Rich World", in Martin Greenberger (ed.), Computers, Communications, and the Public Interest, Baltimore, Johns Hopkins Press, 1971, pp. 37 to 52. Quoted passage verified via Timothy Taylor, "Economics of Information Overload: Thoughts from Herb Simon", Conversable Economist, 17 August 2015. conversableeconomist.com/2015/08/17/economics-of-information-overload-thoughts-from-herb-simon/

• Penn State Digital Shred, "Designing organizations for an information-rich world (Simon)", 7 May 2021. Second source for the citation and for the opening sentence of the passage, "In an information-rich world, the wealth of information means a dearth of something else". sites.psu.edu/digitalshred/2021/05/07/designing-organizations-for-an-information-rich-world-simon/

• Stanford Institute for Human-Centered Artificial Intelligence, The 2025 AI Index Report, Stanford University, 2025. Source for the fall in inference cost from $20.00 to $0.07 per million tokens at GPT-3.5-equivalent MMLU performance, the "more than 280-fold reduction", the rise in organisational AI use from 55% to 78%, and the rise in generative AI use in at least one business function from 33% to 71%. hai.stanford.edu/assets/files/hai_ai_index_report_2025.pdf

• Stack Overflow, 2025 Developer Survey, AI section. Source for 84% using or planning to use AI tools (up from 76%), 32.7% trusting output accuracy against 45.7% distrusting, 66% citing "AI solutions that are almost right, but not quite", 45.2% citing debugging time, and the fall in positive sentiment to 60%. survey.stackoverflow.co/2025/ai

• UX Tools, Design Tools Survey, AI Adoption section. 2,220 respondents, data collected November 2024 to January 2025. Source for 77.6% adoption among individual contributors, 85.2% among leaders, 88.7% among agency design leaders, 78.2% among corporate leaders, and "75.2% Percentage of AI usage focused on text-based tasks". uxtools.co/survey/ai-adoption/overview

• United States Copyright Office, Copyright and Artificial Intelligence Part 2: Copyrightability, A Report of the Register of Copyrights, January 2025. Source for the findings that prompts alone do not provide sufficient human control, that creative selection, coordination or arrangement of AI-generated material is protectable, and that existing law is adequate. copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf

• AIGA, "AIGA Standard Form of Agreement for Design Services". Source for the five-module structure with Basic Terms and Conditions and Intellectual Property Provisions required, the medium-specific modules optional, and the quoted statements about custom proposal documents. aiga.org/resources/aiga-standard-form-of-agreement-for-design-services

• Alina Valyaeva, "AI Image Statistics", Everypixel Journal, 15 August 2023. Source for the estimate of more than 15 billion text-to-image images, 34 million per day, and the per-platform breakdown, together with the extrapolation methodology that makes it an estimate rather than a measurement. journal.everypixel.com/ai-image-statistics

Related reading

• Latent Space as a Material

• Simondon, and Why Good Tools Converge

• Provenance, Disclosure, and the Mechanical Turk

• Dark Patterns Stopped Being a Style Question

More from the Moc Mallen Studio journal