O conteúdo está em inglês; a navegação e os filtros mantêm o idioma selecionado.

“AI Is Really Too Expensive”: The Available Evidence Does Not Support the Broad Claim

The candidate consists almost entirely of a broad opinion about AI being expensive, without naming a model, provider, price, or event. Most supplied sources address hallucinations, jobs, or writing style rather than pricing.

Publicado em 06/08/2026, 09:18:54Última atualização 06/08/2026, 09:18:54Publicação editorial

The candidate does not present a verifiable factual narrative. Its title and excerpt offer an impression rather than details such as a product name, billing period, API rates, subscription terms, or a comparison standard. It is therefore impossible to determine whether the claim concerns inference fees, enterprise subscriptions, hardware, or a specific policy change.

The evidence is also poorly matched to the topic. The BBC material concerns fabricated information appearing in search and chatbot responses, while the other sources discuss employment, AI-generated prose, and industry expectations. None directly establishes that a particular AI service is overpriced.

A 2025 Stanford AI Index report points in the opposite direction for a broad industry claim: the cost of inference at comparable capability levels fell sharply from 2022 to 2024. That trend does not mean every premium model is cheap, but it conflicts with an unqualified claim that AI as a whole is becoming too expensive. The claim should therefore be treated as unverified.

Evidências das fontes 6

A verificação de fontes não garante exatidão permanente. Considere a fonte original, a data e as correções posteriores.

介绍2025年人工智能指数报告 - Stanford HAIhai.stanford.edu · Complementar

价格衡量。人工智能词元定价数据来自 Artiﬁcial Analysis 和 Epoch AI 的 API 定价专有数据库，报告价格为输入与输出词元价格的 3:1 加权平均值为分析推理成本。人工智能指数与 Epoch 合作测量了在固定人工智能性能阈值下成本的下降情况。这种标准化方法有助于更精确的比较。虽然新模型价格可能更高，但其性能也显著提升——若直接与性能较低的老模型比较，可能掩盖真实趋势：单位美元获得的人工智能性能已大幅提高。例如，在 MMLU 测试（评估语言模型性能的常用比较基准）中达到 GPT-3.5 水平（64.8 分）的模型，其推理成本从 2022 年 11 月的每百万词元 20 美元降至 2024 年 10 月的 0.07 美元（Gemini-1.5-Flash-8B），约 1.5 年内下降超 280 倍。在 GPQA （比 MMLU 更具挑战性的比较基准）上得分超过 50% 的模型的成本也呈现出类似的趋势。在这方面，推理成本从 2024 年 5 月的每百万词元 15 美元降至同年 12 月的 0.12 美元（Phi 4）。 Epoch AI 估计，根据任务不同，大语言模型的推理成本正以每年 9 至 900 倍的速度下降。 2022–2024 年选定的比较基准的推理成本资料来源: Epoch AI, 2025; Artiﬁcial Analysis, 2025 | 图表：2025 年人工智能指数报告推理成本（美元/ 百万词元- 对数刻度）图 1.3.22 发布日期 GPT-3.5 Llama-3.1-Instruct-8B Gemini-1.5-Flash-8B GPT-4o-2024-05 Phi4 Claude-3.5-Sonnet-2024-06 GPT-4-0314 DeepSeek-V3 [...] 账户中的照片制作而成，导致她感到恐惧、羞耻和焦虑，并对她的社交和学业生活造成了影响。尽管施害者面临少年司法处罚和学校纪律处分，但此案暴露了应对人工智能驱动骚扰的法律和制度框架存在漏洞。贝里及其家人随后呼吁加强保护措施，美国国会已提出多项法案，旨在将未经同意分享亲密图像（真实或虚假）的行为定为犯罪，并要求社交媒体平台履行删除义务。部分国家，包括澳大利亚，已通过