Factual error
A factual claim is incorrect, such as a wrong date, definition or calculation. Check the claim using appropriate original or credible sources, or work through the calculation yourself.
AI Limitations & Fact-Checking Strategies
Learn to recognise AI limitations, spot hallucinations, and build verification into your study workflow.
🚨 Critical Reality Check
AI tools can be useful learning assistants, but fluency is not evidence. They can confidently present false information, fabricate sources, and make mathematical errors. Understanding these limitations is essential for using AI responsibly in your academic work.
Never submit work containing AI-generated facts, statistics, interpretations or citations without verifying them yourself.
Four Kinds of Error to Check
A response can contain more than one kind of error. Checking only whether a reference exists will miss important failures.
Fabricated evidence
A reference, quotation, statistic or other evidence is invented. Locate the original publication or dataset and confirm the quoted or reported material exists.
Misrepresentation
A real source exists but does not support the claim made about it. Read the relevant passage, table or dataset in context and check the scope, qualifications and limitations.
Reasoning error
Facts or sources may be genuine, but the conclusion does not follow adequately from them. Examine assumptions, missing steps, alternative explanations and whether the evidence is sufficient.
Fluency Is Not Evidence
A polished answer can still be wrong. Source grounding, citations, web access and agreement between AI systems can assist investigation and verification, but do not themselves establish truth. Systems may share sources, assumptions or failure patterns.
Use AI to generate possible explanations, questions and checks. Use original sources, data, worked calculations, official guidance and your own judgement to decide what is supported.
Conservative
Directly supported by the source or data you have checked. Lowest risk for academic writing.
Ask: What evidence supports this?
Interpretive
A reasonable reading of the evidence, but it depends on assumptions you should name.
Ask: What additional evidence is needed?
Speculative
Useful for brainstorming, but not suitable as a claim unless you can verify it independently.
Ask: What would count against this?
Essential Fact-Checking Strategies for Students
Verification is structural: build these checks into your prompt, notes and editing process rather than leaving them until the final proofread.
Why: AI models can generate plausible-sounding but incorrect numbers.
How:
- Cross-reference with official sources (government statistics, academic databases)
- Check multiple independent sources
- Look for the original data source, not secondary reports
- Be especially careful with comparative statistics
Example: If AI says "60% of students use AI for homework," find the actual study, check the sample size, date, and methodology before citing it.
Why: AI frequently fabricates citations that sound legitimate but do not exist.
How:
- Search for the exact paper title in Google Scholar
- Verify author names and publication years
- Check if the journal or publisher exists
- Access the actual source to confirm it says what the AI claims
- Use university library databases for academic sources
Warning: Before citing, verify that the source exists and that the original passage supports the specific claim. Relevance alone is insufficient; do not present invented references as evidence.
Why: AI can be confidently wrong about technical details, formulas, and specialised knowledge.
How:
- Check formulas and calculations manually or with specialised software
- Verify technical specifications against manufacturer documentation
- Consult authoritative textbooks or academic sources
- Ask subject matter experts (lecturers, tutors) if uncertain
Tip: For mathematical problems, work through the steps yourself. For scientific facts, check recent peer-reviewed papers.
Why: Different models may expose weak spots, alternative interpretations or missing checks. Agreement may help identify points to investigate, but it is not proof.
How:
- Ask the same question to more than one tool when the issue matters
- Look for differences, uncertainty and assumptions
- Investigate any discrepancies thoroughly
- Investigate suitable source-discovery or source-grounded tools through the Toolkit; check every important claim yourself
Strategy: If models disagree, do additional research using academic databases. If they agree, still verify the claim in a source you can cite.
Why: Training information can be outdated. Supplied documents, retrieval, tools or web search may provide newer information, but an answer can still be wrong.
How:
- Know your AI tool's knowledge cut-off date
- Use tools with web search for recent information
- Verify any "current" information with news sources or official websites
- Be cautious with evolving fields (technology, medicine, current events)
Note: Knowledge cut-off dates and web access change regularly. Check the current help page for the tool you are using, especially for recent events, policy, law, medicine, prices or technical specifications.
Choose Evidence Appropriate to the Claim
Use original research and datasets, relevant course readings, worked calculations, authoritative textbooks or official guidance as appropriate. Check publication dates, methods, scope and limitations. Read enough context to determine what the source actually supports.
Search and discovery services can help locate candidate sources. Their summaries and citations are not substitutes for checking the originals. Find current access links in the Toolkit.
Source-Grounded Tools Help, but They Are Not Magic
Source-grounded systems can work from selected documents or retrieved passages. This can make claims easier to trace and verify and can reduce some forms of unsupported generation. It does not guarantee accurate interpretation. See Gemini Notebook (formerly NotebookLM) for a product-specific example.
However, source-grounded does not mean automatically correct. The tool may miss a relevant passage, over-weight a weak match, misread a quotation, or sound confident from only part of the evidence.
Use It For
- Finding candidate passages in readings you have chosen
- Generating questions for closer reading
- Comparing what a small set of sources appears to say
Still Check
- Does the cited passage actually support the answer?
- Are important sources missing from the notebook?
- Has the tool confused summary with analysis?
Understanding Context Windows & Token Limits
The context window (also called context length) is the maximum amount of text an AI model can process at once. The context available to a model may include instructions, conversation text and selected document passages; it is not necessarily the whole history or every uploaded file.
What Counts Towards the Limit
- Conversation messages included by the system
- Prior AI responses included in the current context
- Source passages selected or retrieved from supplied files
- System instructions (invisible to you)
- The new message you are sending
What Happens at the Limit
- Oldest messages may be "forgotten"
- Context from early conversation is lost
- You may need to start a new conversation
- Uploaded documents might be truncated
- Some tools show warnings before reaching limits
What to Check Before Uploading a Large File
Capacity
How much text can this tool process at once, and does it warn you when material is skipped?
Retrieval
Does the tool search the whole file, selected passages, or only a summary of the upload?
Privacy
Will the upload be stored, reviewed, shared, or used for training? Are you using a personal or institution-approved account?
Verification
Can you trace every important answer back to the original passage, page, table or dataset?
Note: Model limits and product settings change regularly, so check the current documentation for the exact tool you are using.
Best Practices for Managing Context
✅ Do These
- Start new conversations for different topics
- Summarise key points periodically in long conversations
- Upload only relevant sections of long documents
- Be concise in your prompts when possible
- Use tools with larger context for document analysis
- Save important responses outside the conversation
❌ Avoid These
- Keeping one conversation going for weeks
- Uploading entire textbooks when you need one chapter
- Assuming the model remembers everything from the start
- Mixing multiple unrelated topics in one conversation
- Ignoring context limit warnings
- Relying on very old context for current questions
🎯 The Golden Rule of AI Verification
If you cannot verify it, do not cite it. If you cannot explain it, you do not understand it.
You remain responsible for explaining and supporting the claims you submit. AI should support rather than replace required reasoning. Follow assessment-specific permission, disclosure and privacy guidance.