{"id":"47a14a97-a91f-4226-9521-d5014c246164","arxiv_id":"2412.04503","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A readable review of LLM concepts and limitations that argues LLM errors are not hallucinations but confident outputs with no regard for truth.","lead":"This paper is an introductory review of large language models, covering how they are built, adapted, and how they fail. It is mainly useful as a readable overview for people entering the field, rather than as a source of new findings.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'bullshit' argument's load-bearing premise—'no algorithm for truth'—is contradicted by the paper's own citation-verification example, leaving the reframing unsupported.","rationale":"The reader identified the same load-bearing assumption: the 'no algorithm for truth' premise in Section 4.4. I agree that this premise is the weakest point. My stress-test sharpens it by noting an internal inconsistency: the paper itself verifies the nonexistence of fabricated references via links and databases, which is an algorithmic separation of true and false outputs. This is not an external philosophical disagreement; it is an internal contradiction in the argument's support. The conclusion that 'hallucination' is purely subjective loses its stated foundation once this premise fails. I also note a second instability: Frankfurt's concept of bullshit presupposes an agent capable of being indifferent to truth, whereas LLMs lack intentions or beliefs; at minimum the paper needs an explicit operational extension. However, the paper is an expository primer rather than a research claim requiring acceptance, and the reader's UNVERDICTED verdict already reflects that no formal verification applies. The concern would be important if the philosophical claim were the paper's main contribution, but it does not change the appropriate verdict for a primer that is otherwise descriptive. Hence UNCHANGED.","tokens_in":19044,"tokens_out":4534,"duration_ms":50504,"concrete_test":"Run an automated ground-truth verification on a mixed set of LLM outputs, including the paper's own reference-fabrication example: prompt several LLMs for twenty academic references on 'LLM orchestration,' check each against Crossref/DOI and publisher databases, and add simple verifiable-fact queries (arithmetic, capitals, letter counts). If an automated checker separates true from false outputs with high accuracy, the premise that 'there is no algorithm for truth' is falsified, and Section 4.4's argument for 'bullshit' over 'hallucination' needs a different foundation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 4.4—that LLMs 'do not hallucinate, they produce bullshit'—rests on two linked premises: (i) truth evaluation of outputs is impossible because 'there is no algorithm for truth,' so calling an output a hallucination is a post-hoc subjective value judgment; and (ii) LLMs' lack of comprehension makes them Frankfurtian bullshitters. Premise (i) is internally contradicted by the paper's own evidence: the authors report that suggested references 'did not exist,' 'links to the papers did not resolve,' and 'conferences and journals cited do not list the papers.' Those are exactly algorithmic verifications of truth. The general undecidability of truth does not imply that no meaningful separation of outputs is possible; '2+2=4' versus '2+2=5' and existing versus fabricated DOIs are decisively separable. Premise (ii) is also unstable: Frankfurt's bullshit is defined by a speaker's indifference to truth, an intentional state, and a language model has no such state; the paper extends the term without supplying an operational criterion for what 'indifference' means in a system with no beliefs or intentions. These are correctness risks in the paper's strongest conceptual claim, not merely disagreement with philosophical consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a primer on large language models: it surveys transformer architectures (encoder-only, decoder-only, and encoder-decoder), pre-training objectives, fine-tuning and prompt-engineering techniques, orchestration with retrieval and knowledge-representation systems, and a set of risks (catastrophic forgetting, model collapse, jailbreaks, and hallucination). Its most distinctive and emphasized claim is in Section 4.4: LLMs do not hallucinate, they produce bullshit in Frankfurt's technical sense, because they generate with indifference to truth; the paper argues that the term 'hallucination' is a post-hoc human value judgment and that one cannot separate true from false outputs because 'there is no algorithm for truth.' The paper is written for a broad academic and industry audience and relies entirely on cited literature plus a few anecdotal own-experiments.","tokens_in":19273,"tokens_out":3565,"duration_ms":37116,"significance":"As a synthesis, the survey portions have real value for non-specialists: the organizational structure is clear, the figures and tables are helpful, and the treatment of orchestration and mitigation strategies is pragmatic. The paper is also honest about limitations and cites recent, relevant work. Its conceptual contribution is concentrated in Section 4.4, where the 'bullshit' reframing borrows directly from Hannigan et al. and Hicks et al. but adds a disputed premise about the absence of an 'algorithm for truth.' That premise is not defended and is contradicted by the paper's own citation-checking example, so the paper's most distinctive claim currently rests on unstable ground. The factual misattributions in the survey further reduce confidence in a document whose stated purpose is to orient newcomers.","major_comments":[{"comment":"The claim that 'one cannot separate one language generation output from another in a meaningful way because there is no algorithm for truth' is overstated and internally contradicted by the authors' own verification example in the same section. The authors report that suggested academic references 'did not exist,' that 'links to the papers did not resolve,' and that 'conferences and journals cited do not list the papers.' These are precisely algorithmic or quasi-algorithmic checks of factual adequacy. General undecidability of truth (e.g., for arbitrary mathematical statements) does not imply that all factual claims are inseparable; '2+2=4' versus '2+2=5,' or an existing DOI versus a fabricated one, are decidably separable. Since this premise is load-bearing for the argument that 'hallucination' is merely a subjective judgment, the reframing is unsupported as written.","section":"Section 4.4, 'Hallucinations and their Impacts'"},{"comment":"The paper adopts Frankfurt's definition of bullshit as communication by a speaker who is 'indifferent to the truth,' but it does not supply an operational criterion for what indifference means for a system with no beliefs, desires, or intentions. The citation to Hicks et al. does not resolve this difficulty because the paper's own formulation ties the bullshit conclusion to the 'no algorithm for truth' premise, which fails as argued above. Without either a defended account of model-level indifference or a revised definition, the conclusion that 'LLMs do not hallucinate, they produce bullshit' does not follow from the premises given.","section":"Section 4.4, Frankfurtian bullshit definition"},{"comment":"The survey contains several factual errors that are significant for a primer whose purpose is reliable orientation: RoBERTa is attributed to 'researchers at Google' (Section 2.2), 'peer-to-peer (p2p) from Google' is listed as a decoder-only model (Section 2.2.1), and the GPT listing in Section 2 includes 'GPT-1o preview and GPT-1o mini,' which are inconsistent with the actual model names. These errors are easily corrected but, in a document aimed at non-specialists, they materially undermine trust in the survey's accuracy. They should be fixed before publication.","section":"Sections 2.2 and 2.6"}],"minor_comments":[{"comment":"Typos and stylistic slips: 'property view' should be 'properly view' (Section 4.4), 'in tact' should be 'intact' (Section 2.5), 'vasts amounts' should be 'vast amounts' (Section 2.3), 'accomodate' should be 'accommodate' (Section 2.1), and 'lastlycompletion' should be 'lastly, completion' (Section 2.6.1).","section":"Throughout"},{"comment":"The paper inconsistently names models and products, e.g., 'BARD' instead of 'Bard,' and the list of GPT variants (GPT-3, GPT-4, GPT-4o, GPT-1o preview, GPT-1o mini) is internally inconsistent. Please align names with the official product names.","section":"Section 2.6"},{"comment":"The statement that the training data allocation is 'usually set around 15%' appears without a citation and is not generally true; typical splits are larger. Please clarify or remove.","section":"Section 2.5"},{"comment":"The claim that ChatGPT o1-preview 'will correctly answer 3 because it does parse the word and then double check itself' is presented as a definitive fact without a systematic test; given the paper's own emphasis on anecdotal evidence, this should be softened or supported.","section":"Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads more like a white paper or an editorial survey than a standard research article. Its novelty is concentrated in Section 4.4, and that argument currently rests on an undefended and internally contradicted premise about truth. If the journal accepts position pieces or survey-plus-commentary contributions, major revision could make it acceptable; otherwise, the paper may be better placed in a practitioner-oriented venue. The authors should also perform a thorough fact-check of model attributions and names before any resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a primer, not a research paper, and it should be judged as one. It compiles a lot of standard LLM material into an accessible structure, and the sections on fine-tuning, jailbreaks, and model collapse are genuinely useful for a non-specialist. The one idea with any spark—that LLMs don't hallucinate but produce Frankfurtian bullshit—is borrowed from Hannigan et al. and Hicks et al., and the authors push it hard. That's also where the paper gets into trouble.\n\nThe bullshit argument rests on the claim that 'one cannot separate one language generation output from another in a meaningful way because there is no algorithm for truth.' But the authors themselves separate outputs meaningfully a few paragraphs earlier: they report that the references their LLMs suggested 'did not exist' and the links did not resolve. That is exactly an algorithmic, checkable verification. The undecidability of general truth doesn't mean no statement about the world can be checked; 'there is no algorithm for truth' is doing too much work. And Frankfurt's bullshit is defined by the speaker's indifference to truth, which is an intentional state—applying it to a system with no beliefs needs a bridging argument, not a hand-wave. This matters because the paper presents the reframing as a conceptual correction, not just a stylistic preference.\n\nThere are also some plain factual errors that should embarrass a primer: RoBERTa was not developed by Google, there is no well-known Google model called 'p2p', and 'GPT-1o preview' is presumably o1-preview. These are small individually, but a primer needs to be trustworthy on exactly this kind of detail. The cross-model hallucination similarity is reported as an anecdote, not a controlled observation, and it should be labeled as such.\n\nWho is this for? Students and practitioners who need a first map of the LLM landscape. They'll get value from the carefully organized sections, but they should not take the bullshit argument at face value. Would I cite it? Probably not; the existing surveys it draws on are more reliable. Would a serious editor send it out? Yes, if the journal publishes tutorials—a referee can catch the factual slips and push the authors to either defend or soften the philosophical claim. That's a useful editorial role.\n\nIn sum: the scaffolding is good, the load-bearing wall in section 4.4 is cracked, and the paper deserves revision, not rejection out of hand.","headline":"A competent and occasionally useful LLM primer that trips over its own philosophical headline: the claim that LLMs 'produce bullshit' is undercut by the authors' own verifiable example.","tokens_in":19769,"tokens_out":4267,"would_cite":false,"duration_ms":40271,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LLMs do not hallucinate; they produce bullshit in the technical sense, the authors argue.","keywords":["large language models","hallucination","bullshit","prompt engineering","jailbreak attacks","model collapse","catastrophic forgetting","retrieval augmented generation"],"falsifier":"A single reproducible case in which a model's generation demonstrably depends on the truth of the proposition—for example, a model that consistently refuses to state a false claim even when the prompt rewards falsehood—would show that indifference to truth is not total and would undercut the blanket claim that all LLM output is bullshit.","tokens_in":18854,"feed_emoji":"🤖","tokens_out":9357,"duration_ms":82069,"temperature":0.7,"pith_summary":"This paper is a primer on large language models that also advances a conceptual correction. Its central argument is that the term 'hallucination' misdescribes what LLMs do: a model that generates a plausible but false answer is not misperceiving reality but producing language with no regard for whether the claim is true—bullshit, in the technical sense of the word. The authors ground this in the fact that LLMs are trained to predict likely continuations of text rather than to represent a world, and in the claim that labeling an output a hallucination is a subjective human judgment because there is no algorithm for truth. If the argument holds, the practical problem is not to correct the model's perception but to treat every unverified output as untrustworthy, and the paper reviews mitigation strategies accordingly.","feed_headline":"LLM errors are bullshit, not hallucinations, authors argue","feed_subtitle":"A primer reframes the core failure mode: models generate with indifference to truth, so verify before trusting.","key_machinery":"The central object is the concept of bullshit in its technical sense: communication produced with indifference to the truth. The mechanism that earns this label is the decoder's next-token prediction objective, which optimizes for plausible continuation rather than factual correctness. The paper additionally relies on the premise that no algorithm for truth exists, so judging an output to be a hallucination is a subjective human evaluation. This framing does the work of shifting the mitigation question from 'how do we fix the model's perception?' to 'how do we manage a generator that is structurally unconcerned with truth?'","core_discovery":"On the paper's own terms, the central claim is that LLM errors belong to a different category than perceptual errors. Because decoder-based language models are designed to produce the most plausible continuation of a prompt, they are indifferent to the factual status of what they generate; they do not first form a belief and then misstate it. Calling an output a 'hallucination' treats the error as a failure of perception, but the authors argue this is a post-hoc value judgment made by a human, since there is no algorithm for truth against which the output can be checked. The appropriate description, they argue, is bullshit in the technical sense: speech made without concern for the truth, which may accidentally be true or false. The paper supports this by noting that several different LLMs generated similar false references when asked for canonical academic citations, which it attributes to shared architectures and training data.","pith_inferences":["The paper does not spell this out, but the bullshit framing implies that verification should be built into LLM applications: retrieval-augmented generation and tool use are not just enhancements but the primary defense against truth-indifferent generation.","If the 'no algorithm for truth' premise is taken literally, it also rules out any objective benchmark of factual accuracy, which would undercut the very hallucination indexes the paper cites; a more moderate reading would restrict the claim to open-ended generation.","A testable extension would be to measure whether models can be trained to emit calibrated uncertainty or explicit 'I don't know' responses; if such training succeeds, it would show that indifference to truth can be reduced even if not eliminated.","The bullshit framing invites a philosophical worry the paper sets aside: bullshit originally describes an intentional human attitude, and applying it to a statistical model may be metaphorical; the practical conclusions survive that worry, but the name may not."],"forward_implications":["If LLM errors are truth-indifferent rather than perceptual, then no amount of scaling or fine-tuning will make a hallucination-free model; the goal becomes detecting and managing untrustworthy output.","Users should treat LLM output as a starting point, verifying critical claims against trusted sources instead of relying on the model's fluency.","The term 'hallucination' should be retired in technical discussion, because it smuggles in a subjective judgment that the output is false rather than describing the generation process.","Similar false outputs across different LLMs are to be expected, since models share architectures and training data, so cross-model agreement is not evidence of correctness.","Mitigation strategies for related risks—jailbreak attacks, catastrophic forgetting, model collapse—are partial, so continual evaluation of model outputs remains necessary."],"supporting_citations":[{"why":"Supplies the technical definition of bullshit as communication indifferent to truth.","marker":"[15]"},{"why":"Introduces 'botshit' and frames chatbot outputs as bullshit in this technical sense.","marker":"[19]"},{"why":"Argues directly that a major chatbot produces bullshit in the technical sense.","marker":"[21]"},{"why":"Establishes the transformer architecture whose next-token prediction objective is the mechanism behind truth-indifferent generation.","marker":"[67]"}],"fun_headline_variants":["LLM mistakes are bullshit, not hallucinations, says primer","Why LLM errors are bullshit, not hallucinations","LLMs don't hallucinate; they bullshit, paper argues","Reframing LLM errors: bullshit over hallucinations","LLM outputs: bullshit by design, not hallucination"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the premise that there is no algorithm for truth, so calling any output a hallucination is a subjective human judgment rather than a factual description.","fun_headline_variants_meta":{"raw":{"variants":["LLM mistakes are bullshit, not hallucinations, says primer","Why LLM errors are bullshit, not hallucinations","LLMs don't hallucinate; they bullshit, paper argues","Reframing LLM errors: bullshit over hallucinations","LLM outputs: bullshit by design, not hallucination"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000131,"raw_usage":{"total_tokens":1037,"prompt_tokens":764,"completion_tokens":273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":380,"completion_tokens_details":{"reasoning_tokens":189}},"tokens_in":380,"tokens_out":273,"duration_ms":3078,"temperature":1.0,"reasoning_tokens":189,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:49:03.764010+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A single reproducible case in which a model's generation demonstrably depends on the truth of the proposition—for example, a model that consistently refuses to state a false claim even when the prompt rewards falsehood—would show that indifference to truth is not total and would undercut the blanket claim that all LLM output is bullshit.","supporting_citations":[{"cited_title":"Princeton University Press, Princeton","cited_arxiv_id":null,"evidence_quote":"Supplies the technical definition of bullshit as communication indifferent to truth."},{"cited_title":"Hannigan, Ian P","cited_arxiv_id":null,"evidence_quote":"Introduces 'botshit' and frames chatbot outputs as bullshit in this technical sense."}],"review_version":1}