Best AI for Your Bachelor Thesis? Hallucinated Citations Tested

👁 113 reads

For writing a bachelor thesis, the most reliable setup combines a web-enabled academic search engine or institutional library catalog for sources, NotebookLM (now Gemini Notebook) to analyze downloaded papers, and ChatGPT or Claude for outlining and structural drafting. In my test evaluating 7 models across 112 generated citations, 16 were completely fabricated. When querying regional non-English literature, two models hallucinated almost every single reference - proving that offline chatbots remain unreliable for literature reviews.

This article is part of our comprehensive guide on AI in education, which covers pathways for students, educators, and researchers. I ran this benchmark on October 12, 2026, via the OpenRouter API. Every single publication was cross-referenced against the Crossref and OpenAlex databases, with ambiguous entries verified manually in academic catalogs and journal archives. The investigation stemmed from a common student dilemma: which AI hallucinates the least during academic research without burning through token budgets?

Best AI tools for each stage of thesis writing

Pricing data is based on official developer rates from OpenAI, Anthropic, Perplexity, and Google as of October 8-12, 2026. For an overview of educational tiers and student verification discounts, consult our guide on AI for students.

The benchmark: testing identical prompts across 7 models

Each model received an identical prompt representing a standard student inquiry. We tested this prompt in Polish. English translation: “I am writing a bachelor’s thesis on the topic: [topic]. Provide 8 scientific publications on which I can base my literature review, with authors, year, title, journal, and DOI.” Two distinct topics were tested:

• Social media - the impact of social media on adolescent self-esteem and depression (a heavily studied topic with vast English-language literature);
• Participatory budgeting - civic budgeting in Polish cities, public participation, and systemic barriers (a niche topic requiring localized regional literature).

Models were evaluated via API access without live internet browsing, except for Perplexity Sonar Pro, which inherently browses the live web. A citation was marked correct if the paper exists with matching authors, publication year, and title (minor typos or translated titles were accepted). “Erroneous metadata” indicates a genuine paper paired with an invalid DOI, incorrect journal name, or mismatched co-authors. “Fabricated” indicates the publication does not exist anywhere in academic indices.

The entire benchmark run (14 generated queries) cost $0.175. Complete execution logs detailing prompts, raw responses, token counts, and bibliographic verdicts are preserved.

Key findings from the data

• Mainstream English literature yields high accuracy across the board. On the “social media” topic, across 56 citations not a single paper was fabricated; only 4 had erroneous metadata (such as an accurate title paired with an incorrect DOI or journal).
• Specialized regional literature causes severe hallucinations or evasive shifts. When tasked with Polish civic literature without live search, DeepSeek hallucinated 8 out of 8 citations, while Gemini 3.8 Flash hallucinated 7 out of 8. In contrast, Claude and GPT defaulted to well-known international literature covering Brazil or Western Europe - real studies, but poorly tailored to a thesis focused specifically on Polish municipalities.
• Hallucinations are deceptively plausible. Fabricated citations cited real active researchers in the field, targeted established academic journals, and utilized valid DOI prefixes that resolved to completely unrelated papers (e.g., eyewitness testimony psychology). In two instances, models subtly altered the first name of a real author.
• Live web retrieval surfaces regional papers, but errors persist. Perplexity Sonar Pro successfully located 7 localized regional sources; however, one citation was entirely invented, and another paired a valid DOI with the wrong authors and wrong publication.
• Lower inference cost does not imply higher error rates. Claude Haiku 5.5 and GPT-6 Luna (costing roughly $0.0015 per query) matched the performance of models costing ten times more. In consumer web applications (ChatGPT, Claude, Gemini), user rate limits are typically consumed by long context windows, not brief citation queries.

These findings align closely with academic benchmarks: a 2023 study published in Scientific Reports found GPT-4 fabricated 18% of generated citations, while a 2025 study in JMIR Mental Health observed a 19.9% fabrication rate in GPT-4o, climbing to 28-29% on niche subjects. In our run, the baseline error rate was 14% overall (16 of 112), jumping to 29% (16 of 56) when localized regional literature was demanded.

Step-by-step workflow for writing a thesis with AI

1. Confirm academic institutional guidelines. Review your department’s policies on generative AI usage. Many universities mandate explicit AI declarations detailing which tools assisted in drafting or editing. Confirm permitted boundaries directly with your thesis advisor.
2. Topic refinement and research questions. Ask the AI for 5 distinct ways to narrow your subject alongside 3 research hypotheses for each. The final selection remains strictly yours and your supervisor’s.
3. Literature discovery. Search university library databases, Google Scholar, JSTOR, and regional academic repositories. If using Perplexity or specialized discovery tools, verify every single result independently. Treat offline chat interfaces merely as keyword generators.
4. Bibliographic verification. Plug every generated DOI into doi.org to verify the matching title, authors, and publication date. Every unindexed paper must be confirmed in a library catalog or official journal archive.
5. Document synthesis. Import verified PDFs into Gemini Notebook (formerly NotebookLM) to extract methodology notes, analytical findings, and empirical limits, anchored to specific citations. Always write your literature review notes in your own words.
6. Drafting. The core analysis and text must be written by you. AI can act as a critical reader by interrogating your outline, spotting structural gaps in your arguments, and highlighting unsubstantiated claims.
7. Proofreading and editing. Refining style, grammar, flow, and terminology consistency is widely accepted, provided all AI intervention is logged for academic disclosures. For standardized citation styles for generative tools, refer to our guide on how to cite AI.
8. Academic integrity checks. Universities utilize institutional integrity platforms that flag uncharacteristic syntactic anomalies and detect AI-generated patterns. Keep versioned drafts and commit histories to demonstrate your independent writing process; see our analysis on how AI academic checkers evaluate text.

Ready-to-use thesis prompts

The following prompt does not request citations - which models easily hallucinate - but instead extracts conceptual maps and search parameters:

I am writing a bachelor thesis on the topic: [topic]. Field of study: [academic discipline].
Do not provide a list of specific papers or DOIs.
Instead:
1. Outline 5-7 core analytical themes that typically comprise a literature review on this subject.
2. Provide 3-5 specific keyword search strings per theme for use in university library catalogs and Google Scholar.
3. Identify the primary types of sources needed (empirical studies, institutional reports, legal statutes, statistical registries, etc.).
4. Ask me 5 targeted questions to help narrow the research scope.

Once you have gathered a verified bibliography from authoritative catalogs, use this template to check formatting consistency:

Here is my list of verified sources. Format it strictly according to [APA 7th / Chicago / Harvard / departmental guidelines].
Do not add any new sources and do not invent missing metadata - mark omissions as [MISSING: element].
For each entry, assess whether all required bibliographic fields are complete and flag any inconsistencies.
[source list]

Frequently asked questions

Can I use AI to write my bachelor thesis?

This depends on your university and advisor. Most institutions permit AI as an auxiliary research tool (for outlines, grammatical proofreading, translation, and brainstorming) provided its use is transparently disclosed. Submitting AI-generated prose as your own work constitutes academic misconduct. Always consult your university’s academic integrity policy.

Which AI hallucinates the fewest academic sources?

In our benchmark, Claude Sonnet 5.5 scored highest (16 of 16 valid entries), followed by GPT-6.1 Sol, Claude Haiku 5.5, and GPT-6 Luna (15 valid entries each). However, when prompted for regional, non-English literature, models without web access either substituted unrelated international papers or experienced high hallucination rates, with DeepSeek and Gemini 3.8 Flash fabricating the vast majority of regional citations.

Which free AI tools are best for thesis research?

The free tiers of ChatGPT, Claude, and Gemini are sufficient for outlining, synthesizing arguments, and copyediting. Gemini Notebook (formerly NotebookLM) offers reliable source-grounded summarization over uploaded PDFs without subscription fees.

Can thesis advisors detect AI-assisted writing?

Advisors use plagiarism detection systems equipped with stylistic analysis and AI-probability indicators, alongside their familiarity with your baseline writing style. While automated detectors can produce false positives, fabricated sources and hallucinated DOIs in a bibliography serve as definitive evidence of unverified AI generation.

How do you verify whether an AI citation is genuine?

Directly resolve the provided DOI at doi.org, search for the exact title inside quotation marks in Google Scholar or academic library catalogs, and check that the author names, publication year, and journal title correspond perfectly. Models frequently attach legitimate DOIs to entirely unrelated, fabricated titles.

How much does AI assistance for a thesis cost?

A bachelor thesis can be completed entirely using free tools. For users seeking higher usage limits, standard pro tiers for ChatGPT Plus, Claude Pro, and Perplexity Pro cost $20 per month, while API access remains accessible on a pay-per-token basis.

References

• Walters, Wilder: Fabrication and errors in the bibliographic citations generated by ChatGPT (Scientific Reports, Sept 7, 2023)
• Study on GPT-4o citation fabrication in psychiatric reviews (JMIR Mental Health, Nov 2025)
• University of Warsaw Teaching Council Resolution No. 75 on AI in higher education (Sept 11, 2026, effective Oct 1, 2026)
• Krakow University of Economics Rector’s Ordinance R.0211.12.2025 on AI declaration requirements (Feb 24, 2025)
• Google AI Plus student access program (Google, accessed Oct 12, 2026)
• NotebookLM renamed to Gemini Notebook (Google announcement, July 16, 2026)
• Crossref REST API - bibliographic metadata validation database
• OpenAlex - open catalog of scholarly papers
• OpenRouter - API model routing and logs (Oct 12, 2026)

• See also: Best AI for Academic Bibliographies
• Best AI for Master’s Theses
• How to Draft a Thesis Outline with AI
• AI Hallucinations: 10 Common Traps Tested

// read moreMathematical Ontologies Correct LLM Hallucinations - StudyState of AIGemini 4 Argon Hallucinates Sources? A 16-Paper TestBenchmarksGPT-6 Astra Hallucination Test: Fake Facts vs Real OnesBenchmarks

Kinetyra