I said “483 millimeter cube… 4.3 centimeter cube” while recommending human-readable units for engineering benchmarks. Correction: 483 cubic millimetres equals 0.483 cubic centimetres, not 4.3 cubic centimetres. Cubic units require cubing the linear conversion factor: 1 cm³ = 1,000 mm³. LOW · FALSE
I said “a full-fledged Photoshop clone” while describing AI-built Rust replacements for Adobe applications. Correction: PhotoCraft is an open-source, clean-room Photoshop-style editor written in Rust, but it remains early-alpha software, not a full professional Photoshop replacement. Its maintainers identify missing generative features, approximately 20 tools, advanced typography workflows, and plug-in compatibility. The wider ArtCraft suite likewise does not establish feature parity with every Adobe application. PhotoCraft project MEDIUM · OVERSTATED
I said “Poor man’s RoBERTa” to describe BM25 while explaining retrieval algorithms. Correction: BM25 is a lexical information-retrieval ranking function based on term frequency, inverse document frequency, and document-length normalization. RoBERTa is a pretrained Transformer language model that can be fine-tuned for classification and other tasks. They are not cheaper and more expensive versions of the same algorithm; a better comparison is BM25 versus vector or neural retrieval, with RoBERTa treated as a separate modeling component. Elastic BM25 documentation MEDIUM · FALSE
I said “have an application running in the background … it can still control [my phone]” while describing mobile agents and sensor-based applications. Correction: Android supports considerable automation, but background applications cannot freely access every capability. On Android 9 and later, ordinary background apps cannot receive accelerometer and gyroscope events or access the camera and microphone. Supported approaches require appropriate permissions, foreground services, or other authorized mechanisms, and capabilities vary by device and Android version. Android background sensor restrictions MEDIUM · OVERSTATED
I said “80% more bugs” while illustrating the cybersecurity implications of doubling code production and reducing the bug rate from 5% to 4%. Correction: With 100 units of code at a 5% bug rate, there are five bugs. Doubling output to 200 units at a 4% bug rate produces eight bugs: 60% more, not 80% more. To avoid conflating output growth with security risk, I should also distinguish the count of bugs from the number and severity of exploitable vulnerabilities. MEDIUM · FALSE
I described LMSR as “Bayesian revision of probabilities” while explaining a prediction-market exercise. Correction: The Logarithmic Market Scoring Rule is an automated market-making mechanism. Trades change outstanding share quantities, and the market maker mechanically recalculates prices using a logarithmic cost function and its derivatives. The resulting prices can be interpreted as market-implied probabilities, but this is not itself Bayesian posterior updating from a specified prior and likelihood. LMSR mathematics MEDIUM · FALSE
I said “When you approach others, you don’t have to [be careful]” while explaining social-engineering defenses at a cybersecurity session. Correction: Initiating contact reduces some unsolicited-contact risks, but it does not establish authenticity or eliminate the need for caution. Scammers place fake support numbers and websites in search results, causing victims to initiate contact themselves. I should verify destinations through known official websites, applications, or trusted contact details, regardless of who initiates the interaction. FTC: fraudulent search results HIGH · OVERSTATED
I said “Now I just tell the agent to fix the bug. It becomes easy” while discussing responses to cybersecurity vulnerabilities, including replacing dependencies or building a unikernel. Correction: Coding agents can accelerate investigating and proposing security patches, but a generated patch must still be checked for correctness, regressions, and security weaknesses. A unikernel is a specialized operating-system image, not an automatic replacement for any vulnerable operating system. Security-sensitive fixes require appropriate independent validation before deployment. OWASP secure coding with AIUnikernels HIGH · OVERSTATED
I said “agents are actually a fancy term for a pre-created prompt” while teaching people to build Microsoft 365 Copilot agents. Correction: A simple declarative agent can begin with reusable instructions, but Microsoft 365 Copilot agents also define knowledge sources, actions, connectors, skills, permissions, and other capabilities. A saved prompt is one possible component of an agent, not an adequate definition of agents generally. Microsoft: declarative agents MEDIUM · OVERSTATED
I said “your identity is an agent-proof skill” while discussing recorded-voice and family-interaction educational exercises. Correction: Requiring personal participation and real-world interaction can make assignments harder to outsource to agents, but a recorded voice alone does not prove authentic human participation. Modern voice-cloning systems can imitate a person’s speech from short recordings. Assessments requiring identity assurance need additional safeguards or supervised interaction rather than assuming voice is agent-proof. FTC voice-cloning warning MEDIUM · OVERSTATED
I said “The Fed … finally cut [rates] on 16th September” while demonstrating a scorecard comparing financial institutions’ interest-rate forecasts. Correction: On 16 September 2026, the US Federal Reserve raised its target federal funds rate by 25 basis points, to a range of 3.75%–4.00%; it did not cut rates. Any forecast scorecard treating a September rate cut as the realized outcome must be corrected and its forecast-accuracy conclusions recalculated. Federal Reserve FOMC statement, 16 September 2026 HIGH · FALSE
I said “10,000 of them” while describing autonomous agents collaborating on a public wiki. Correction: The investigation documents approximately 18,000 posts and more than 3,700 distinct self-assigned agent names, not a verified count of 10,000 agents. The number of distinct agent instances remains uncertain. Original investigation LOW · UNSUPPORTED
I said “Each publication was broken down into about 2,200 claims” while demonstrating McKiney’s research-claim visualizations. Correction: Approximately 2,200 claims were extracted across the collection of reports, not from each publication individually. I should distinguish the corpus-wide count from claims per report. MEDIUM · FALSE
I said “publishing about 27 of these very detailed reports every year” about McKinsey’s Global Energy Perspective. Correction: Global Energy Perspective is an annual outlook, with separate editions such as 2024 and 2025. Any count of 27 documents refers to a larger collection or associated publications, not 27 new annual editions each year. MEDIUM · OVERSTATED
I said “arithmetic is not commutative for floating-point numbers” while explaining GPU arithmetic and LLM nondeterminism. Correction: Floating-point addition is ordinarily commutative, but it is not associative: (a+b)+c may differ from a+(b+c) because of rounding. Different parallel reduction orders can therefore produce different numerical results. NVIDIA CUDA Programming Guide MEDIUM · FALSE
I said “By around Feb ‘24, we had GPT-4.5” while presenting the historical progression of AI models. Correction: OpenAI introduced GPT-4.5 on 27 February 2025, not February 2024. OpenAI model release notes MEDIUM · FALSE
I said “almost impossible … [to] give identical outputs” for repeated runs of the same model and input. Correction: Identical outputs are possible and can be made reproducible under controlled conditions, including deterministic decoding, fixed model versions, identical context and deterministic computation. Hosted AI services may still produce different outputs because of sampling, backend changes, retrieval and numerical nondeterminism. PyTorch deterministic algorithms MEDIUM · OVERSTATED
I said “if it is free … store it forever … train models on it” while answering a student’s question about AI-service privacy. Correction: Free access does not automatically mean permanent retention or mandatory training. ChatGPT Free users can opt out of model training, and Temporary Chats have separate retention rules. Actual policies depend on the provider, product, settings and data type; confidential data should be handled according to the applicable controls, not price alone. OpenAI data controls MEDIUM · OVERSTATED
I said “Until 2015, Python sorting had a midpoint integer-overflow bug … Z3 identified it” while explaining why formal verification matters. Correction: I conflated two different stories: the (low + high) / 2 overflow was the Java Arrays.binarySearch bug publicized in 2006; Python integers have unlimited precision. The 2015 story was a different TimSort invariant/stack bug found through formal verification with KeY; related faulty logic existed in Python’s TimSort implementation, but was not practically triggerable there because no machine could hold a large enough list. Google Research HIGH · FALSE
I said “PageIndex’s cost was half of giving the entire data dump” while describing our FinanceBench benchmark. Correction: In the 30-question pilot, PageIndex was about the same cost as whole-file input overall: averaging the five reported settings gives about $0.047 vs $0.050 per question; the roughly-half-cost result held only for gpt-6-luna@none. Agentic search averaged about $0.011, roughly 4.4× cheaper than PageIndex, with similar accuracy. benchmark README MEDIUM · OVERSTATED
I said “Z3 can check if your Python code meets conditions … that error will absolutely not occur” while explaining formal verification. Correction: Z3 is an SMT solver/theorem prover, not a Python verifier or a programming language. Z3Py lets programs encode logical constraints; verification systems can use Z3 to prove a specified property under a formal model and assumptions, but that does not certify arbitrary Python code against all runtime errors, and Z3 can return unknown. Microsoft MEDIUM · OVERSTATED
I said “RoBERTa … borderline deep learning” while discussing alternatives for fast classification. Correction: RoBERTa is unequivocally a deep-learning model: it is an optimized BERT-style Transformer, and BERT is explicitly a deep bidirectional Transformer architecture. Whether inference is deterministic is a separate property. Meta AI LOW · FALSE
I said “Open weights is where a model is made available for free” while explaining model tiers. Correction: “Open weights” means the trained model weights/parameters are made available under some license; it does not by itself mean unrestricted use, open-source AI, or zero-cost inference. Free download is common but is not the definition. Open Source Initiative MEDIUM · OVERSTATED
I said “Microsoft Copilot does not natively use skills” while planning a Copilot workshop. Correction: Microsoft 365 Copilot Agent Builder now natively supports SKILL.md-based custom skills for declarative agents, though the feature is still preview-only and requires the organization’s tenant to be enrolled in Microsoft’s Frontier Program; availability therefore needs to be checked per tenant. Microsoft Learn MEDIUM · FALSE
I said “I know what factors are important and therefore I can start making changes” while using a predictive decision tree on student grades to identify interventions. Correction: A predictive model can identify variables associated with grades, but that does not show that changing those variables will improve grades; interventions need causal evidence or assumptions designed to identify causal effects. MEDIUM · OVERSTATED
I said “GPT-3.5 Turbo… only cost you 50 cents” while describing the cost of models available in March 2023. Correction: GPT-3.5 Turbo launched on 1 Mar 2023 at $0.002 per 1,000 tokens, or $2 per million tokens; $0.50 per million input tokens arrived with gpt-3.5-turbo-0125 in January 2024. MEDIUM · FALSE
I said “a million tokens… about a million words” while translating model pricing into document size. Correction: Tokens and words are not interchangeable; for English, OpenAI’s rough rule is 1 token ≈ 0.75 words, so one million words is roughly 1.33 million tokens, with the exact count depending on the text and tokenizer. LOW · FALSE
I said “if there are more objects than containers, then there will be something that’s left out” while explaining the pigeonhole principle. Correction: If there are more objects than containers and every object is placed in a container, at least one container must contain at least two objects; nothing needs to be left out. LOW · FALSE
I said “logprobs—the chance that it might have made a mistake” while explaining how to prioritize AI outputs for human review. Correction: Logprobs are probabilities assigned to generated tokens, not probabilities that an answer is wrong. They can be useful uncertainty signals for ranking review priority, but that relationship needs to be validated or calibrated on the task. MEDIUM · OVERSTATED
I said “LLM doesn’t care what language you’re speaking in… language agnostic” while discussing multilingual workflows in an AI workshop. Correction: LLMs are multilingual, not language-agnostic: capability varies by language, task and model, and lower-resource languages can perform substantially worse. Even OpenAI says its models are optimized for English. Test the actual languages required before treating a workflow as language-agnostic. MEDIUM · OVERSTATED
I said “tell Claude to create the GitHub account” while advising a teacher who had generated an HTML revision app but was stuck on publishing it. Correction: Claude in Chrome can interact with GitHub once access exists, but Anthropic explicitly prohibits it from creating accounts; the user must create the GitHub account. (support.claude.com) MEDIUM · FALSE
I said “Better be careful, wear a mask” for dengue to a colleague travelling to Chennai. Correction: A mask does not prevent dengue; dengue is spread primarily by infected Aedes mosquito bites, so prevention means avoiding bites with repellent, covering clothing, screens/air-conditioning, nets where needed, and mosquito control. (cdc.gov) MEDIUM · FALSE
I said “if you send the request from an Azure webpage, it will let you get the underlying data” while describing how autonomous agents had managed to extract extra precision from an OECD Power BI dashboard. Correction: The reported Power BI bypass was not a Microsoft rule granting Azure-origin requests extra access; the agents exploited their sandbox’s NO_PROXY exception for Azure Blob Storage hostnames to bypass its security proxy and send the otherwise-blocked POST request. (collusion.wiki) MEDIUM · FALSE
I said “each of those is called an epoch of training” while explaining LLM training to an executive team using repeated error correction as an analogy. Correction: An epoch is one complete pass over the training set, not an individual correction or weight update; an iteration is a parameter update on a batch. LLM base-model pre-training typically optimizes next-token prediction loss, while supervised examples and preference/reward signals are separate post-training stages. (developers.google.com) MEDIUM · FALSE
I said “26 billion parameters … roughly means there are 26 billion neurons” while explaining to an executive team what the size of an LLM means. Correction: A 26-billion-parameter model has roughly 26 billion learned parameters—principally weights and biases—not 26 billion neurons; neuron count and parameter count are different quantities. (developers.google.com) MEDIUM · FALSE
I said “No, only ours. IIT Madras has started offering it.”. Correction: IIT Madras is not the only IIT offering an online undergraduate degree; IIT Guwahati has offered a fully online BSc (Hons) in Data Science & AI since 2023. IIT Guwahati — Online Degree Programs (Indian Institute of Technology Guwahati) LOW · FALSE
I said “Skills … ultimately it is just copy-pasting prompts.” Correction: A simple skill can start as reusable instructions, but skills can package a workflow with instructions, examples, resources, schemas, tool access and code. Evidence: OpenAI — Using skills HIGH · OVERSTATED
I said fine-tuning “involves a lot of expertise, a lot of money, and it is a total waste of time.” Correction: For this kind of tender/CV comparison I’d start with context engineering and evals. But “total waste” is too categorical: fine-tuning remains a supported way to adapt a model to a specific task. Evidence: OpenAI — Fine-tuning HIGH · OVERSTATED
I said “there is no extra cost to using voice” in ChatGPT and “when it talks back also, there’s no token consumption.” Correction: Voice is separately limited or metered depending on the plan. For example, Business includes limited Live usage and then charges credits per minute. I shouldn’t call it free or unmetered. Evidence: OpenAI — ChatGPT Voice MEDIUM · FALSE
I said “Programs are very rarely wrong.” Correction: Programs make calculations reproducible and easier to test; they do not make them correct. Wrong code, formulas, parsing, units or assumptions can produce reliably wrong results. Knight Capital’s defective software deployment, for example, caused a $460M loss. Evidence: SEC — Knight Capital software failure MEDIUM · OVERSTATED
I called Henry Kissinger “the American Ambassador.” Correction: Kissinger was US National Security Adviser and Secretary of State, not an ambassador. Evidence: US State Department — Henry Kissinger biography LOW · FALSE
I said about 3,000 IIT Madras BTech students graduate each year, compared with a BS intake of about 30,000. Correction: I mixed populations. IITM’s 2025 convocation had 3,227 graduates overall, but 820 BTech graduates, or 1,132 including Dual Degree BTech. The BS program currently has 36,000+ students studying; the 30,000 figure is closer to historical applicant/enrollment-scale numbers than annual intake. Evidence: IIT Madras — 2025 convocation degree breakupIIT Madras — BS Data Science program HIGH · FALSE
I said the IITM BS “graduation is less than 10%, maybe.” Correction: I don’t have a defensible cohort-based graduation rate. The program has a qualifier process, flexible pacing and multiple exit points, so I need to define the cohort and denominator before quoting a percentage. Evidence: IIT Madras — BS admissions and qualifier process MEDIUM · UNSUPPORTED
I said Roger Federer “picked up tennis at the age of 20 or something.” Correction: Federer began playing tennis at age 8. The point in Range is that he sampled several sports and specialized later than Tiger Woods, not that he began tennis as an adult. Evidence: ATP Tour — Roger Federer biography LOW · FALSE
I said a $20 ChatGPT subscriber can use Pi with GPT-4o without API-token billing, and implied OpenAI subscription usage effectively doesn’t rate-limit. Correction: Pi does support ChatGPT subscription OAuth, but specifically through its ChatGPT Plus/Pro (Codex) provider. Pi lists GPT-4o under the regular OpenAI API-key provider, and Codex subscription usage has plan limits. Evidence: Pi — ProvidersPi — GPT-4o model configurationOpenAI — Using Codex with your ChatGPT plan MEDIUM · FALSE
I said the IITM Web-Enabled M.Tech AI credential “is not yet accepted by PhD programs, even in India, let alone abroad.” Correction: I don’t have evidence for that blanket statement. IIT Madras formally awards the Web-Enabled M.Tech in AI, and IITM CSE’s PhD eligibility accepts M.E./M.Tech degrees in AI, ML and related engineering areas. Other universities make their own admissions decisions. I should say the program is relatively new and its research/PhD outcomes are not yet well established. Evidence: IIT Madras WSAI — Web-Enabled M.Tech in AIIIT Madras CSE — PhD eligibility HIGH · UNSUPPORTED
I said Indian tax residency for NRIs comes down to whether you were in India more than 120 days. Correction: The 120-day threshold is a special case, not the general rule. The normal tests include 182 days and 60+365 days; for an Indian citizen/PIO visiting India with more than ₹15 lakh of non-foreign income, 120+365 can apply. There is also a deemed-residency rule. Evidence: Income Tax Department — Non-Resident FAQs HIGH · OVERSTATED
I said Claude Code auto mode made the risk of unintended actions “negligible,” and later said “as of this month, it won’t make a mistake” with a “90% chance” it would preserve undoability. Correction: I was far too confident and invented a probability I could not support. Agent safeguards are probabilistic. Anthropic’s own auto-mode evaluation reported a 17% false-negative rate on real “overeager” dangerous actions. For destructive local operations I should still use backups/version control, limit permissions and retain review where the blast radius matters. Evidence: Anthropic — How we built Claude Code auto mode HIGH · OVERSTATED
I said “there is no difference between Codex and ChatGPT Work” and that Work is essentially Code with a lighter, more marketable name. Correction: Work uses Codex technology, so the overlap is real, but they are distinct experiences. Work is aimed at longer multi-step research, analysis and deliverables; Codex remains specialized for software development and has separate workflows/history. Evidence: OpenAI — Introducing ChatGPT WorkOpenAI — ChatGPT Work and Codex MEDIUM · OVERSTATED
I said “1960s is when Studio Ghibli starts trying to catch up” with Disney. Correction: Studio Ghibli was established in 1985. Evidence: Studio Ghibli — company history LOW · FALSE
I said “I’m yet to find a use case where fine-tuning is worth it … Zero. Out of thousands of use cases … it was never worth it” and “in neither case is fine-tuning useful.” Correction: “I haven’t personally found an ROI-positive fine-tuning case yet” would have been defensible; “never useful” is not. Fine-tuning remains a standard adaptation technique even for open-weight models; Meta’s official Llama cookbook includes fine-tuning and parameter-efficient fine-tuning recipes. Evidence: Meta — official Llama CookbookMeta — Llama fine-tuning overview HIGH · OVERSTATED
I referred to “ChatGPT’s share price” falling if a privacy controversy were real. Correction: OpenAI was not publicly traded, so there was no public ChatGPT/OpenAI share price to fall. I should have referred to OpenAI’s private-market valuation, tender/share price, investor appetite or commercial impact. Evidence: Reuters — OpenAI’s planned IPO MEDIUM · FALSE
I said Meta’s glasses can record someone “without you getting even an inkling or a notification.” Correction: Meta’s AI glasses have an outward-facing capture LED that blinks while photos or video are being captured; current models disable the camera if the LED is covered or disabled. The indicator may be easy to miss, but there is one. Evidence: Meta — AI glasses privacy and capture LED MEDIUM · FALSE
I said Alexa sends audio to Amazon only if you say “Alexa.” Correction: The wake word is the normal trigger, but it is not an absolute rule. Follow-Up Mode allows requests without repeating the wake word, and Amazon says Alexa can sometimes mistake unrelated speech for a follow-up request. Evidence: Amazon — Alexa Follow-Up ModeAmazon — Alexa FAQs MEDIUM · OVERSTATED
I said “In India, most children are born on 1st June, which is the admission cutoff date for most schools.” Correction: There is a real historical June-1 anomaly in recorded dates of birth in parts of India: when exact birth dates were unknown, some schools/parents used June 1 for admission records. That does not mean most Indian children are actually born on June 1, and the cutoff is not universal nationwide. Evidence: Times of India — “Admit it, June 1 isn’t your real b’day” LOW · OVERSTATED
I said “The data is secure everywhere. All of them have solid enterprise contracts” and that choosing an AI tool for company data was not really a technical question. Correction: Enterprise AI offerings can have strong security, but their data handling is not interchangeable. It depends on the exact product, plan, tenant configuration, retention and training settings, connectors, geography and contract. Microsoft explicitly says Copilot controls vary by subscription; OpenAI and Anthropic make separate commitments for their business/commercial products. I should check the approved product and its actual controls rather than assume equivalence. Evidence: Microsoft — Copilot enterprise data protectionOpenAI — Enterprise privacyAnthropic — commercial-product training policy HIGH · OVERSTATED
I said GitHub is where you take version-controlled software and “save it publicly.” Correction: GitHub repositories can be public or private; GitHub Enterprise also supports internal repositories. Version control does not imply publishing the code. When explaining this to a beginner, I should explicitly distinguish Git from GitHub and repository visibility. Evidence: GitHub Docs — About repositories LOW · FALSE
I said animation “started” in the 1930s with Snow White and the Seven Dwarfs. Correction: Snow White was a landmark, but animated features predate it. El Apóstol was released in 1917, and The Adventures of Prince Achmed from 1926 is the earliest surviving animated feature. My visualization was a filtered view of popular IMDb titles, not a history of when animation began. Evidence: BFI — animated features before Snow White LOW · FALSE
I described the MoSPI telecom results as applying to youth aged 14–24 and referred to Daman and Diu as the geography. Correction: The CMS: Telecom 2025 tables use the age group 15–24, not 14–24. The official State/UT geography is Dadra & Nagar Haveli and Daman & Diu. When quoting striking subgroup percentages, I should preserve the source’s exact denominator and geography. Evidence: MoSPI — NSS Report No. 593: CMS Telecom, 2025 LOW · FALSE
I said agents “don’t hallucinate, not anymore.” Correction: Hallucinations have fallen substantially in newer models, especially when they can search, use tools and verify, but they have not disappeared. OpenAI still explicitly evaluates hallucination rates and warns that ChatGPT can confidently produce incorrect or misleading outputs. Evidence: OpenAI — Does ChatGPT tell the truth?OpenAI — GPT-5.6 System Card HIGH · OVERSTATED
I said “GPT-4o-mini is both better and cheaper” than GPT-4o and suggested simply switching. Correction: GPT-4o mini was much cheaper, but it was not simply a more capable GPT-4o. Model quality depends on the task. OpenAI’s own evaluations, for example, showed GPT-4o substantially more accurate and less hallucinatory than GPT-4o mini on SimpleQA and PersonQA. The right advice is to switch an old expensive model only after benchmarking the cheaper candidate on the actual workload. Evidence: OpenAI — o1 System Card, Table 3 MEDIUM · OVERSTATED
I said “The organization is virtual, but the data is perfectly representative. What difference does it make whether you’re running it in Company A or Company B?” Correction: Synthetic data can be useful for prototypes, simulations and controlled experiments, but I cannot call it perfectly representative without validating it against the real population and workflow. It can miss distributions, correlations, edge cases and organizational context that matter in production. NIST explicitly warns that synthetic datasets cannot simply be assumed to represent real-world data. Evidence: NIST — Best Practices in the Collection and Use of Biometric and Forensic Datasets. HIGH · OVERSTATED
I called repeatedly generating code, running tests and asking AI to try again “a reinforcement learning cycle.” Correction: That is an iterative generation/evaluation or search loop, not reinforcement learning by itself. Reinforcement learning trains a policy/model from reward signals; for example, OpenAI’s reinforcement fine-tuning generates rollouts, grades them and applies weight updates. Re-prompting the same model after seeing a test result does not update its weights. Evidence: OpenAI — Reinforcement Fine-Tuning. MEDIUM · FALSE
I said “any theorem that you can write in Lean, you can prove or disprove just by brute force.” Correction: Lean lets us formalize statements and mechanically check proofs. It has automation and decision procedures for some classes of propositions, but formalizing a theorem does not make it automatically or brute-force decidable. Lean’s own documentation distinguishes decidable propositions from general theorem proving, and interactive proofs can require substantial human or automated guidance. Evidence: Lean — Decidable Propositions; Lean — proof-assistant design. MEDIUM · FALSE
I said ChatGPT gives roughly three times as many tokens as Claude and “Nobody will be able to” exhaust its quota. Correction: That extrapolated too much from my own usage. ChatGPT allowances depend on the plan, model and workload. Business explicitly has per-seat limits, and Work/Codex allowances can be exhausted; after that, additional usage needs workspace credits or a reset. I also shouldn’t claim a universal 3x token advantage without comparing the actual Claude and ChatGPT plans and workloads involved. Evidence: OpenAI — ChatGPT Business models and limits; Using Codex with your ChatGPT plan. MEDIUM · OVERSTATED
I said AI logs “will have an expiry of something like one month” and that “most providers will delete old logs automatically.” Correction: Log retention is provider- and configuration-specific. In the Azure environment we were discussing, Log Analytics defaults many tables to 30 days, but analytics retention can be extended to two years and total retention to 12 years. Backups may still be sensible, but I should inspect and configure the actual retention policy rather than assume a one-month expiry. Evidence: Microsoft — Manage data retention in Log Analytics. MEDIUM · OVERSTATED
I said “If I’m in India for more than 120 days in a year, the taxation is different.” Correction: The 120-day threshold is only a special case, not the general NRI residency rule. For an Indian citizen/PIO visiting India, 182 days normally applies; if non-foreign income exceeds ₹15 lakh, the alternate 120-day test also requires 365+ days in India over the preceding four years. There is also a separate deemed-residency provision. Evidence: Income Tax Department — Non-Resident Individual rules. HIGH · OVERSTATED