AI Glossary / AI and machine learning / Intermediate
Self-correction
Definition
Self-correction is a system's ability to detect problems in its own output and revise it, either within one run through reflection and checking, or across runs through feedback loops.
What it means
Modern AI systems make mistakes with confidence, so the interesting engineering question is whether the system can catch its own. Self-correction builds that in: generate, critique, revise.
Within a single run, a model can be asked to check its work against constraints, verify arithmetic, or re-read a source it claims to quote. Across runs, user ratings and downstream outcomes feed back so tomorrow's attempt avoids today's error.
The honest limitation: a model checking itself shares the same blind spots as the model that erred. Self-correction works best when the check is cheaper or sharper than the generation, like running code, querying a database, or matching against rules.
How it works
Generate a first attempt at the answer or plan.
Run explicit checks: constraint lists, calculators, retrieval of quoted facts, test cases for code.
Feed any failures back to the model with instructions to revise, and repeat a bounded number of times.
Across runs, collect ratings and outcomes, and use them to adjust prompts, retrieval or training.
Measure the correction loop itself: how often revisions actually fix errors, and how often they introduce new ones.
Cap revision rounds and route unresolved outputs to a person rather than looping indefinitely.
Three practical examples
- A drafting assistant that re-reads the source contract and verifies each clause it cited before showing the summary.
- A coding agent that runs its own code, reads the error message, and patches the failing line before reporting done.
- A support bot whose resolved-versus-escalated outcomes tune next week's reply templates.
Think of it this way
Think of proofreading with a checklist. The writer misses their own typos, but a checklist of known traps, names, dates, totals, catches what fresh eyes would. Self-correction is the checklist, not a second brain.
Worked example
the unchecked total
Suppose an AI assistant summarizes a supplier proposal with payment terms. The figures are illustrative.
Its first draft states a discount of fifteen per cent; a checking step re-reads the proposal and finds the fifteen applies only to year one, with five thereafter.
The revision corrects the summary and, just as important, flags that the first draft dropped a condition, a pattern worth watching.
The general lesson: the correction is valuable, but the log of what needed correcting is where the system-level improvement lives.
Illustrative case study
A logistics firm let an agent draft carrier booking emails, then added a self-check: before sending internally for approval, the agent re-read the rate confirmation and verified every figure and date in its draft. Caught errors fell by four fifths, and the team noticed the remaining ones clustered around handwritten amendments, which routed those to humans by rule. The firm's conclusion: self-correction is not about trusting the AI more, it is about giving it the same checking tools a careful human would use, and measuring what it still misses.
Common misconceptions and mistakes
- Self-correction makes outputs reliable. It reduces error rates, sometimes sharply, but never to zero; high-stakes outputs still need human review.
- Reflection alone fixes reasoning. Asking a model to think again helps mainly when paired with external checks like tools, data or rules.
- More correction rounds are better. Each round costs time and can drift the answer; two or three bounded rounds is the usual design.
- Self-correction replaces human review. It reduces the volume reaching humans and sharpens what remains; the final call on high-stakes outputs stays with people.
Business use
Design self-correction around external checks, tools, rules and sources, not just the model re-reading itself.
Bound the loop: set a maximum number of revision rounds and escalate after that.
Log correction events; clusters of corrections point at fixes worth making upstream.
Keep humans for the errors self-correction provably misses, identified from the logs.
Track the correction rate as a health metric; a rising rate means the upstream prompt, model or data is degrading and needs attention.
Start with the two or three checks that catch your most expensive error type, prove them in production, and only then widen the loop's remit.
People also ask
What is self-correction in AI systems?
The ability to detect problems in its own output and revise, through in-run checking loops or across-run feedback from outcomes.
Why is external checking better than reflection alone?
The same model that made the error shares its blind spots; tools, rules and source data catch errors the model cannot see in itself.
Where is self-correction most valuable?
Wherever checks are cheap and objective: arithmetic, code execution, quoted facts, format constraints and policy rules.
Does self-correction slow systems down?
Yes, each check adds latency and cost, so loops are placed only where errors are expensive and checks are cheap, such as figures, quotes and code.
Check your understanding
What is self-correction, and what makes it actually work?
Answer: Self-correction is a system catching and revising its own errors, in-run via critique loops or across runs via feedback. It works when the check is external and sharper than the generator, calculators, source documents, tests and rules, because a model reflecting alone shares its own blind spots; and it works when the loop is bounded, logged and measured, so corrections reduce errors instead of churning them.
Related terms
Learn next
AI Agents and Automation for Small Business covers building exactly these check-and-revise loops into workflows: https://aiforbusinessandlife.com/academy/ai-agents-automation-small-business [C4]
More in AI and machine learning
- Scaling - Scaling is rescaling numeric features to comparable ranges, so that no variable dominates a model merely because of the units it happens to be measured in.
- Schema-agnostic databases - Schema-agnostic databases, in the query-interface sense, let users request information without first knowing the database's exact structure or vocabulary.
- Scoring - Scoring is applying a trained model to new cases to produce its outputs: a probability, a prediction, a risk grade, the number the model was built to give...
- Self-healing test automation - Self-healing test automation is test software that attempts to recover from certain changes or failures in its execution setup.
- Self-management (computer science) - Self-management in computer science is the ability of a computing system to monitor and adjust aspects of its own operation under defined objectives and...
- Self-supervised learning - Self-supervised learning is training where the data itself provides the labels: the model is asked to predict a hidden or rearranged part of the input, so...
Sources
[S2] Google, Machine Learning Glossary; [S5] Anthropic, Building effective agents. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Self-attention Next: Self-healing system