Two terms, defined here so they can be quoted, checked and argued with. Both name something the model-versus-model comparison genre skips: the product wrapped around the model, and the bets its team made about you. I coined both on August 11, 2026. This page is the definition of record.
Perceived AI intelligence
Perceived AI intelligence is the capability we attribute to the model after the surrounding product has supplied context, selected tools, organized the work and presented the result.
Ask which model is best and the honest answer is: it depends. On the job, on your taste, and on how you reach it. A chat interface and a raw API call can run identical weights and feel like different products, because most of what we read as intelligence arrives from somewhere else.
Four layers do that work. Context, which the product assembles from history, files and memory. Tool selection, which decides what the model can reach. Orchestration, which sequences the steps. Presentation, which decides what we see while it happens and what we get at the end.
The consequence is testable. Memory alone can produce a startling increase in perceived intelligence with no corresponding leap in the underlying model. If you have watched the same model behave better inside one product than another, you have already measured this.
What it is not. There is an established research construct called perceived intelligence, with validated measurement scales in human-robot interaction and assistant UX. See Krening and Feigh on characteristics that influence perceived intelligence in AI design (2018) and Ling and colleagues on scale development for AI travel assistants (2025).
That work measures the judgment. This term explains what produces it, and puts the cause in the product layer rather than in the model. Use both words every time. The bare phrase belongs to the researchers above.
Agent barks
Agent barks are the short status lines an AI agent emits while it works, borrowed from the game-writing term for the one-line callouts non-player characters shout to create the illusion of coordination.
A bark is a real and much older term from game writing: the short lines characters yell in reaction to what is happening around them. The word is not mine. Pointing it at AI agent status lines is.
F.E.A.R. shipped in 2005 with soldiers that ran on a planning system called GOAP. They looked coordinated, like a trained unit. They were not talking to each other, because nothing in 2005 could do that. The designers built the illusion instead, under one rule documented in their own postmortem, Combat Dialogue in F.E.A.R.: The Illusion of Communication: if the AI did not say it, it did not happen.
So a soldier yells "Flanking!" and your brain assembles a squad. Game AI designers have known for years that barks make games look smarter. Now open any agentic AI product and watch the status line. Same mechanism, different room.
Barks add no capability. They change how the wait feels. When passengers at a Houston airport kept complaining about baggage waits, the fix that finally worked was moving the arrival gates farther from baggage claim. Longer walk, identical wait, complaints near zero. A streaming reasoning trace is the walk to baggage claim.
| Product | Bark style | What it optimises for |
|---|---|---|
| ChatGPT Work | Barely barks | Quiet surface, least narration |
| Perplexity Computer | Narrates in full sentences | Legibility of work in progress |
| Claude Cowork | A verb plus a running second count | Progress and the cost of waiting |
| Claude Code | Whimsical cosmetic lines, 186 in the shipped binary | Entertainment, not status reporting |
The failure mode. A bark that narrates work which is not happening is verification theater, not legibility.
The five assumptions underneath both terms
Both terms come out of the same comparison. ChatGPT Work, Claude Cowork, and Perplexity Computer share plenty of capabilities, but they start from different assumptions about how we work.
| Assumption | ChatGPT Work | Claude Cowork | Perplexity Computer |
|---|---|---|---|
| We want to be understood | Infers shorthand from history | Infers it inside your project | Infers it across sessions |
| We want it to remember us | Some context reaches new projects | Remembers the room you built | Brain builds the graph in the background |
| We do not like the wait | Barely barks | Verb plus seconds | Full sentences |
| We want more than the model to be smart | Company knowledge from the org | You supply files and folder | Skills can deploy agents |
| We do not like starting over | Absorbs context from chats and apps | Carries context inside the project | Corrections are an explicit input |
Assumption four has its own precedent, and it is also from a game. The Sims shipped in 2000 with characters that were deliberately not very bright. Will Wright's team moved the intelligence into smart objects instead. The fridge announced that it satisfied hunger. The shower signalled hygiene. A Sim never had to understand the world, because the furniture told it what everything was for.
Skills, connectors, tools and MCP servers are the furniture. The model is the Sim. We are the player. Where the three products differ is who carries the furniture up the stairs.
Questions people ask
What is perceived AI intelligence?
The capability we attribute to the model after the surrounding product has supplied context, selected tools, organized the work and presented the result. It names the gap between what a model can do and what a product delivers. Coined by Karo Zieminski in Product with Attitude on August 11, 2026.
Is it the same as "perceived intelligence" in academic research?
No. That literature measures how intelligent users judge a system to be, using validated rating scales. This term explains what produces the judgment, and locates the cause in the product layer rather than in the model.
What are agent barks?
The short status lines an agent emits while working, such as "Searching the web" or "Reading 12 files". Borrowed from game writing, where a bark is a one-line callout a character shouts in reaction to events.
Where does the term "bark" come from?
Game writing, decades before AI agents. F.E.A.R. used barks in 2005 so its GOAP soldiers appeared to coordinate, under the rule: if the AI did not say it, it did not happen. Applying the word to agent status lines is the new part.
Do agent barks make an agent smarter?
No. They change perceived AI intelligence, not capability. Narration makes the wait legible and occupied, and occupied time feels shorter than unoccupied time.
Who coined these terms?
Karo Zieminski, AI product manager and author of Product with Attitude. First use of both: August 11, 2026, in I Build AI Products for a Living. Here Are 5 Things All Model Comparisons Miss.
How to cite these terms
Both definitions are published under CC BY 4.0. Quote them freely, with attribution.
Zieminski, Karo. "Perceived AI Intelligence and Agent Barks: Canonical Definitions." Product with Attitude, August 11, 2026. https://productwithattitude.com/perceived-ai-intelligence-agent-barks.html
Original essay: Zieminski, Karo. "I Build AI Products for a Living. Here Are 5 Things All Model Comparisons Miss." Product with Attitude, August 11, 2026. Read it on Substack.