AI Product Judgment
Your Users Are Already Grading Your AI
Most AI features ship with a thumbs up and a thumbs down button under each answer. In practice the button is optional, so most outputs never get a vote, and the ones that do come from people with a reason to click.
Meanwhile users do something after every single output. They copy it, edit it, ask for another version, or close the tab. Each action is a judgment made by the person the output was written for. Logged, they become a cheap eval running in production, as long as you read them for what they are.
Explicit and implicit feedback
Eugene Yan splits user feedback into two kinds. Explicit feedback is what users say on purpose: thumbs, star ratings, comments. Implicit feedback is inferred from what they do. (Eugene Yan, Patterns for Building LLM-based Systems)
His examples show how much a product already knows. In a coding assistant like Copilot, accepting a suggestion as is counts as a strong positive, accepting it and then editing counts as a positive, and ignoring it is neutral or negative. In Midjourney, rerolling a set of images is a negative signal, asking for variations is a positive one, and upscaling or downloading an image is a strong positive. (Eugene Yan)
Chip Huyen's AI Engineering lists the equivalents for chat products: the user ending the conversation early, correcting the model, the sentiment of their replies, asking for the answer to be regenerated, and how varied the conversation is. She notes that chat makes feedback easy to give but hard to extract. (Chip Huyen, AI Engineering ch. 10; chapter notes by Alex Strick van Linschoten)
The Copilot evidence
The strongest case for these signals comes from GitHub. A study of Copilot users looked at which usage measures best predicted how productive developers felt. The winner was the simple acceptance rate of suggestions. It did better than other measures, including whether the accepted code was still there later. The authors present it as a coarse signal for monitoring, not a full measure of quality: it says how users feel, not which suggestions were wrong or why. (Ziegler et al., Productivity Assessment of Neural Code Completion)
Different actions, different failures
Lumping every negative action into one "dissatisfied" count throws away most of the information.
Picture an AI email-drafting feature with three buttons under each draft: regenerate, shorten, and make more formal. A user who clicks regenerate is saying the draft missed. Maybe it misread the request or got a fact wrong. A user who clicks shorten is saying the content was fine but there was too much of it. A user who asks for a more formal version liked the substance and disliked the tone.
Those are three different failures with three different fixes. A high shorten rate points at a length default in the prompt, not at accuracy. Edits carry even more detail. Research presented at ICML 2024 treats a user's edit to model text as feedback in its own right: the edited version doesn't need to be perfect, it only needs to be better than what the model wrote. (Tucker et al., Coactive Learning for LLMs using Implicit User Feedback)
Even length of engagement is ambiguous. Yan points out that a long conversation could mean the user is engaged, or that they are struggling to get what they need. (Eugene Yan)
Overrides are free labels
One action deserves special attention. When an AI makes a decision on the user's behalf, such as picking a category, a priority, or a routing, and the user changes it, that change is a human label.
The user saw the same input as the model and said what the right answer was. Store the model's guess next to the final value, and every row where they differ becomes raw material for an eval set for that classifier, labeled by the people who know the answer best. In a writing tool I built, the content type the rules guessed is stored next to the type the user finally picked for exactly this reason.
Noise and bias
The case against treating these signals as a score is just as strong.
A regenerate is not always a failure. Some users click it out of curiosity, to see what else the model can do. Huyen lists several biases in user feedback: people are more likely to react to a bad experience than a good one, the users who give feedback are not a random sample, and the system's own behavior shapes the feedback it later receives, which can lock in a loop. (Chip Huyen, AI Engineering ch. 10) A 2025 study of human-LLM conversations found implicit feedback useful for understanding users but noisy as a signal to learn from. (arXiv 2507.23158)
So the experts land in two places at once. Copilot's acceptance rate did track how users felt about the product. (Ziegler et al.) Most practitioners still treat these signals as a way to choose which traces to read, not a number to push up. Hamel Husain and Shreya Shankar list user feedback as one way to sample production traces for review, next to random samples and outliers, and they say to always keep some random traces in the mix. (AI Evals FAQ) Optimize the regenerate rate directly and you might just make the regenerate button harder to find.
Where this fits with other evals
Anthropic describes production monitoring as the layer that shows real user behavior at scale, while being reactive, noisy, and missing ground truth. Its advice is to combine automated evals, production monitoring, and periodic human review, because no single layer catches everything. (Anthropic, Demystifying evals for AI agents)
Behavioral signals sit in the middle. Offline evals check known failures before a change ships, user signals show where new ones appear after it ships, and human review decides what is actually wrong. Husain suggests repeating error analysis on 100 or more fresh production traces every two to four weeks. (AI Evals FAQ) User signals are a good way to decide which traces make that pile.
The loop also runs the other way. Aman Khan suggests comparing eval scores with real user outcomes to find where the evals themselves are wrong. (Aman Khan, Beyond vibe checks) If the offline suite passes an output that users keep regenerating, the suite is missing something.
A practical starting point
If you are adding an AI feature, log every action a user takes on its output, each as its own event, and store the model's original output next to whatever the user ends up keeping. Each week, pull a sample of the outputs with the most regenerates, edits, and overrides, plus a few random ones, and read them. Write down what went wrong, turn the repeat failures into test cases, and resist putting the regenerate rate on a dashboard as a goal.