What is AI Hallucinations about Bogus summer reading list?

That “AI Hallucinations about Bogus summer reading list” refers to a real incident in May 2025 where newspapers published an AI-generated summer reading guide that included books that don’t exist.

*What happened*

- The _Chicago Sun-Times_ and _The Philadelphia Inquirer_ ran a special section called “Heat Index: Your Guide to the Best of Summer.” It included a “Summer reading list for 2025” with 15 books.

- More than half were fake. Only 5 of the 15 titles were real.

- The list was created by freelancer Marco Buscaglia, who admitted he used AI to help research but didn’t double-check the output. He called it “a really stupid error on my part”.

- The content was licensed from King Features, a unit of Hearst, and distributed as a syndicated insert. King Features later fired the writer. 75798be90d842d1f

*Examples of the fake books*

AI “hallucinated” titles and descriptions that sounded plausible, attributing them to real authors:

- _Tidewater Dreams_ by Isabel Allende – described as a “climate fiction novel” about a family confronting rising sea levels

- _The Rainmakers_ by Percival Everett

- _Nightshade Market_ by Min Jin Lee

- _Boiling Point_ by Rebecca Makkai

- _The Last Algorithm_ by Andy Weir – ironically about an AI system that developed consciousness 0d842d1fa0164bc1

*Why this is an “AI hallucination”*

In AI, a hallucination is when a model generates confident-sounding info that isn’t grounded in reality. Large language models can fabricate citations, books, or facts because they predict text patterns, not verify truth. ChatGPT, for example, has invented book titles like _Dynamic Canonicity: A Model for Biblical and Theological Interpretation_ by Harold Coward when prompted. 5c1fa067

*Fallout*

- Chicago Public Media CEO Melissa Bell called it “unacceptable” and said the insert went in without editorial review.

- The Sun-Times Guild said they were “horrified by this slop syndication” and noted it didn’t involve staff reporters.

- It sparked broader concern about newsrooms using AI without fact-checking. a0164bc18be9

*Bigger picture*

Even top models in 2026 still hallucinate. On grounded summarization tasks, frontier models have ∼3% hallucination rates, and on complex docs it can exceed 10%. Researchers consider it a structural issue, not just a bug. a067

So the “bogus summer reading list” became a case study in why AI outputs need human oversight, especially in journalism. a016

What are real examples of a math hallucination and how to catch it?

Here are 3 real patterns of math hallucinations that keep showing up, plus exactly how to catch them:

*1. The fake citation / invented theorem*

*What the AI said*:  

> "By the _Chen-Zhang Theorem (2019)_, any integer > 2 can be written as the sum of a prime and a power of 2. Therefore 7 = 3 + 2²."

*Why it’s a hallucination*: There is no Chen-Zhang Theorem from 2019. The statement it’s “proving” is actually related to Goldbach’s conjecture and is still unsolved. The AI invented both the theorem and the citation to sound authoritative.

*How to catch it*: 

- *Google the exact theorem name + year*: If only the AI result shows up, it’s fake.

- *Check arXiv/Google Scholar*: Real theorems have papers. Search `arXiv Chen Zhang 2019 prime power 2`.

- *Ask for a link*: “Give me the DOI or arXiv ID.” Hallucinations usually reply with broken links or “I made an error.”

*2. The confident wrong arithmetic*

*What the AI said*:

> "To integrate, $\int_0^1 x^2 dx = \left[\frac{x^3}{3}\right]_0^1 = \frac{1^3}{3} - \frac{0^3}{3} = \frac{1}{3}$. Now multiply by 9: $9 \times \frac{1}{3} = 2.7$."

*Why it’s a hallucination*: $9 \times \frac{1}{3} = 3$, not 2.7. The AI did the calculus right but slipped on grade-school arithmetic, then doubled down with steps.

*How to catch it*:

1. *Sanity check simple ops*: Round numbers. $9 \times 0.33...$ should be ∼3, not 2.7.

2. *Re-run in a calculator*: Paste `9 * (1/3)` into your phone. 

3. *Force code output*: Prompt “Solve this and show Python code to verify.” If the code gives 3.0, the text was wrong.

*3. The broken proof with hidden divide-by-zero*

*What the AI said*:

> "Let $a = b$.  

> Then $a^2 = ab$  

> $a^2 - b^2 = ab - b^2$  

> $(a-b)(a+b) = b(a-b)$  

> Divide both sides by $(a-b)$: $a + b = b$  

> Since $a = b$, then $2b = b$, so $2 = 1$."

*Why it’s a hallucination*: Step 4 divides by $(a-b)$, but since $a = b$, that’s $(b-b) = 0$. You can’t divide by zero. The AI presented it as valid algebra.

*How to catch it*:

1. *Check every division*: Ask “What are you dividing by and can it be zero?” 

2. *Plug in numbers*: Let $a=5, b=5$. The step $(a-b)(a+b) = b(a-b)$ becomes $0 \times 10 = 5 \times 0$ which is true. But dividing gives $10 = 5$, false.

3. *Use a proof assistant*: Tools like Lean or Wolfram won’t let that step pass.

*General “catch it” checklist*

Red flag Quick test

**Named theorem you’ve never heard of** Google `"Theorem name" math` with quotes

**arXiv ID or DOI given** Paste it into http://arxiv.org/abs/ID or http://doi.org/ID

**Arithmetic in the middle of a proof** Recalculate that line by hand/code

**“It follows trivially that...”** That’s where steps get skipped. Ask “show the trivial step”

**Too-perfect answer** If it solves an open problem like Riemann Hypothesis, it’s 99.999% hallucinated

*Best defense*: Treat AI like a student showing work. Don’t grade the final answer — grade each line. For anything that matters, run it through Wolfram Alpha, SymPy, or just a calculator.

What is AI Hallucinations in Mathematics?

*AI hallucinations in mathematics* are when a language model or AI tool gives you math that looks correct but is actually made up, wrong, or logically incoherent.

Unlike a normal calculation error, a hallucination is _confidently false_. The AI will show steps, cite theorems, and use notation that seems legit, even though the answer is nonsense.

*Why math is a special problem for LLMs*

LLMs predict text, not truth. They’re trained to sound like a math proof, not actually do the proof. So you get:

1. *Pattern-matching, not calculation*: Ask “What is 1847 × 2931?” and it might guess based on similar problems it saw, not actually multiply.

2. *Fake citations*: “By the Riemann-Zeta Theorem of 2014...” — that theorem doesn’t exist.

3. *Invalid steps that look valid*: Skips logical gaps, divides by zero, or applies a formula where the assumptions don’t hold, but writes it smoothly.

*Common types of math hallucinations*

Type What it looks like Example

**Fake theorems/results** Invents a named theorem or paper “By Gupta’s Lemma (2021), every prime > 5 is twin”

**Wrong arithmetic** Simple calc errors, confidently stated $127 \times 16 = 2,032$ — actually 2,032 is right, but AI might say 2,132

**Broken proofs** Steps don’t connect logically Proves 1=2 by hiding a division-by-zero step

**Phantom references** Cites arXiv papers that don’t exist “See Smith et al., arXiv:2304.12345 for proof”

**Misapplied formulas** Uses correct formula in wrong context Uses L’Hôpital’s rule on a limit that isn’t 0/0

*Why it’s dangerous*

- *Research*: Researchers have submitted papers with AI-generated “proofs” that fell apart under review. 

- *Education*: Students get homework help that’s wrong but looks right, so they don’t catch the error.

- *Engineering/finance*: If you trust an AI for load calcs or pricing models, a hallucinated formula = real-world failure.

*How to spot & prevent them*

1. *Never trust, always verify*: Re-derive critical steps by hand or in a symbolic tool like Wolfram, SymPy, or MATLAB.

2. *Ask for code, not just text*: “Solve this and give me Python to check it” forces the AI to produce executable logic, which fails if wrong.

3. *Use math-specialized tools*: Wolfram Alpha, Lean, Coq, or GPT-4 + Wolfram plugin actually compute. Base LLMs don’t.

4. *Check citations*: If it gives a theorem or paper, search it. Hallucinated arXiv IDs are a dead giveaway.

5. *Make it show constraints*: Prompt “List every assumption you’re using” — hallucinations often violate hidden assumptions.

*Rule of thumb*: LLMs are fluent in the _language_ of math, not the _logic_ of math. They’re great for brainstorming approaches or explaining known concepts, but terrible for “trust me” calculations or novel proofs.

How to AI- proof your contract drafts?

You can't stop AI from hallucinating, but you can build a process so the hallucinations never make it into the final contract draft. Think of it like a QA pipeline.

*1. Control what the AI touches*

- *Segment the work*: Let AI handle low-risk parts like formatting, boilerplate, or first-draft language. Keep it away from facts, numbers, dates, citations, and past performance.

- *No CUI in public tools*: If your draft has Controlled Unclassified Information, don’t use ChatGPT, Claude, etc. Use an on-prem or FedRAMP-approved GovCloud instance.

- *Ground the model*: Give it the exact docs to pull from. “Using only FAR Part 15 and the attached SOW, draft Section L.” Then it’s less likely to invent clauses.

*2. Force source-backing for anything factual*

Make this a rule: _If a fact can be checked, the AI must cite where it came from._

- *Citations required*: “For every case, statute, or FAR clause you cite, include the reporter and a URL to http://govinfo.gov or http://acquisition.gov.”

- *No numbers without source*: Dollar amounts, timelines, percentages need a bracketed source: [Price Proposal tab B-2, cell F14].

- *Ban invented examples*: Tell it: “Do not create hypothetical vendors, awards, or case names. If none exist, say ‘none found’.”

*3. Build a human verification layer*

Treat AI output like a first-year intern’s work: helpful but untrustworthy.

- *Clause checker*: Manually verify every FAR/DFARS/Agency clause number against http://acquisition.gov. AI loves to invent 52.244-99.

- *Case law check*: Run every case through Google Scholar or Westlaw. Hallucinated cases are the #1 issue in legal filings now.

- *Math & dates*: Recalculate all totals, option periods, and delivery dates. LLMs are bad at arithmetic.

- *SAM.gov / CPARS cross-check*: If the AI mentions past performance, pull the actual CPARS report.

*4. Use tooling to catch fabrications*

- *AI-detection + fact-check*: Tools like http://Originality.ai, Pangram, or Harvey can flag text that looks AI-generated. Then you know where to look hardest.

- *Redline comparison*: Diff the AI draft vs your source docs. Anything not traceable gets flagged.

- *Clause libraries*: Use a real clause database like Icertis or Cobblestone, not whatever the AI remembers.

*5. Document your process for compliance*

Agencies are starting to ask. 2024 OMB memo M-24-10 and DoD’s Responsible AI guidance basically say: you used AI, you own it.

- *AI-use log*: Track which sections used AI, which model, which prompts, who verified.

- *Attestation*: Add a cover note: “Sections 1-3 drafted with, verified by on. All facts checked against sources.”

- *Internal policy*: If you’re a contractor, write a policy: “No AI for cost/pricing data. All legal citations must be Shepardized by counsel.”[Tool][Name][Date]

*Quick prompt hygiene that helps*

When you do use AI, prompt defensively:

> “Draft the evaluation criteria section. Use only FAR 15.304 and the attached RFP. If you cannot find support for a statement, write [VERIFY]. Do not invent case names, statistics, or clause numbers.”

*The rule of thumb*: AI can give you 70% speed, but the last 30% is human verification. In government contracts, that 30% is what keeps you out of False Claims Act trouble.

What is AI fabrications in a government contract report?

*AI fabrications in a government contract report* are false, made-up, or misleading statements that an AI tool inserted into a report about government contracts. People also call these “AI hallucinations.” 

How it happens

1. *LLM makes stuff up*: Generative AI tools predict text. If the model doesn’t have the right data, it will still produce confident-sounding sentences that are wrong.

2. *Misapplied to contracts*: When drafting or summarizing a government contract report, the AI might invent case law citations, contract clauses, dollar amounts, dates, vendor names, or past performance data that don’t exist.

3. *Inherits bad inputs*: If the prompt includes errors, or the AI was trained on flawed public data, those errors get amplified in the report.

Why it matters for government contracts

- *Legal risk*: Federal Acquisition Regulation (FAR) requires accuracy. Submitting false statements to the government can violate 18 U.S.C. §1001 – false statements. Even if unintentional, agencies may see it as negligence.

- *Protest risk*: Competitors can challenge an award if the winning proposal or evaluation report contained AI-generated falsehoods.

- *Compliance*: DoD, GSA, and OMB have 2024-2025 memos saying contractors must disclose AI use and are responsible for all content. “The AI did it” isn’t a defense.

- *Audit trail*: Inspectors General now run AI-detection on contract files. Fabricated citations or metrics are a red flag.

Common examples

- *Fake case law*: AI cites _United States v. Acme Corp, 123 F.4th 456_ to justify a cost principle. The case doesn’t exist.

- *Invented past performance*: “Vendor completed $40M bridge for TxDOT in 2023” when no such project occurred.

- *Wrong clause numbers*: References FAR 52.244-99, but that clause isn’t real.

- *Phantom data*: Claims “98.7% of similar IDIQs were protested” with no source.

How agencies & contractors are addressing it

- *Human-in-the-loop review*: Require attorneys/COs to verify every fact, citation, and number before submission.

- *AI use disclosure*: Some RFPs now ask if/where AI was used in proposal prep.

- *Banned tools*: Certain agencies prohibit using public LLMs for Controlled Unclassified Information.

- *Validation*: Run reports through citation checkers, contract clause databases, and http://SAM.gov verification.

Bottom line: If AI writes your contract report, you own every word. Fabrications = false claims risk. Always verify.

Answer of IELTS question:Describe a time when you were really proud of yourself.

*IELTS Speaking Part 2 sample answer: "Describe a time when you were really proud of yourself"*

*You should say:*  

- when it was  

- what you did  

- why you felt proud  

- and explain how it affected you afterwards  

*Sample answer (Band 7-8):*

One time I felt really proud of myself was about a year ago, when I completed my first 10k run. 

I’d never been athletic, but I decided to challenge myself and signed up for a charity run in my city. For three months, I trained almost every morning before work. It was tough at first — I could barely run 1 kilometer without stopping. But I stuck to my plan, gradually increased my distance, and even changed my diet.

On the day of the race, I was nervous but excited. Crossing the finish line in under an hour was such a rush. I felt proud because I proved to myself that discipline and consistency really pay off. It wasn’t just about fitness; I realized I could achieve things that seemed impossible if I put my mind to them.

After that, I became more confident at work and in personal goals. It taught me that progress matters more than perfection.

*Useful vocab/phrases for this topic:*  

- *took on a challenge* = accepted something difficult  

- *stick to my plan* = keep following it  

- *a sense of achievement* = feeling of success  

- *paid off* = gave good results  

- *pushed me out of my comfort zone*  

*Part 3 follow-up questions you might get:*  

1. Is it important for people to feel proud of themselves?  

2. Do you think children should be praised often?  

Answer of IELTS question:What do you do in your free time?

*IELTS Speaking Part 1 sample answer:*

In my free time, I usually read and go for walks. I love getting lost in a good novel because it helps me relax after work. On weekends, I also enjoy cooking new recipes with my family. It’s a nice way to unwind and spend time together. 

*Tips for a band 7+ answer:*

1. *Be specific*: Instead of "I watch TV", say "I watch documentaries about history". 

2. *Add a reason*: Explain _why_ you like it. 

3. *Use range*: Mix simple and complex sentences, add vocab like "unwind", "get lost in", "catch up with".