Retrospective field entries / 120 entries
Platform releases, deployments, incident records, evaluation methods, disclosure rules and the older precedents behind chatbots and agents, each read for the conversation, the system boundary and where it fails.
Historical event dates and source dates are separate from the preparation date of this local edition. Every entry is a retrospective draft prepared 16 September 2026; none was published on its historical date.
120 entries
OpenAI's own announcement describes the plugin manifest, staged rollout and first partners that let ChatGPT browse, retrieve and act.
OpenAI's own help center set a two-stage retirement date and named GPT adoption as its stated reason.
OpenAI's DevDay announcement and help center describe GPT actions and knowledge, and today's version of that page announces its own retirement.
OpenAI's own announcement and current guide describe a persistent voice connection to GPT-4o, distinct from its Assistants API.
Google's announcement and the protocol's own site show A2A moved from a vendor launch to Linux Foundation governance.
The protocol's own changelog lists what changed since March 2025, and its license shows governance moving beyond Anthropic alone.
Anthropic's own engineering essay defines five workflow patterns and reserves the word agent for a narrower case.
AWS's own preview announcement lists AgentCore's components, and its current FAQ still calls some of them preview.
Google's announcement and the ADK repository describe a permissively licensed toolkit distinct from the protocol launched alongside it.
OpenAI's own repository README called Swarm educational and now points every reader to its production successor.
Its March 2023 version reports AlfWorld and HotPotQA gains with GPT-3.5, not the GPT-4 coding numbers added later.
Tree of Thoughts reports GPT-4 gains on three puzzle tasks by letting the model branch, score and backtrack.
Voyager reports large relative gains over its own chosen baselines, plus documented hallucination and cost limits.
Gorilla reports beating GPT-4 on its own 1,645-call API benchmark, a result later folded into a running leaderboard.
WebArena's own 2023 baseline put its best GPT-4 agent at a 10.59% end-to-end task success rate on 812 tasks.
SWE-agent's own paper reports a 12.5% SWE-bench resolve rate with GPT-4 Turbo behind a purpose-built editor.
The framework's own FAQ calls it standalone, splitting autonomous Crews from developer-controlled Flows.
Its own GAIA leaderboard result used GPT-4o, and the framework was later renamed from transformers.agents to smolagents.
OpenDevin's own paper reports a 26% SWE-bench Lite resolve rate with Claude 3.5 Sonnet before renaming to OpenHands.
SIMA's own report says human experts scored only 60% on its hardest test, showing the evaluation's own difficulty.
Turing's own paper proposes a judged game about fooling an interrogator, not a scientific finding about machine thought.
Weizenbaum's 1966 paper explains ELIZA's keyword-matching script and warns against reading comprehension into its replies.
MIT's own reports tie SHRDLU's grammar and inference to a simulated blocks world, not general English dialogue.
RFC 439 records PARRY and ELIZA's DOCTOR exchanging scripted replies in 1972, not a conversation either side understood.
Carpenter's own pages describe Jabberwacky as learning only from stored user conversation, with no authored rules.
The contest's own rules define a win as a relative yearly ranking, and a linked critique says the restricted format proves little.
A.L.I.C.E.'s own AIML documentation describes pattern-to-template matching across tens of thousands of authored categories.
ActiveBuddy's own materials describe SmarterChild as a scripted query agent added to an AIM buddy list, later reaching more networks.
IBM's own account frames Watson's 2011 win as a confidence-scored, closed-domain answer pipeline, not a dialogue system.
Amazon's original Echo page describes a wake-word speaker with a fixed initial skill set, not the assistant Alexa later became.
Google's Dialogflow traces to API.ai, the startup it acquired in 2016, and now splits simple and complex agents into ES and CX.
Rasa's own documentation and GitHub README describe the classic open-source framework moving toward a newer product called Hello Rasa.
Microsoft's 2016 Bot Framework launch and its current Azure documentation show what a channel-agnostic bot actually depends on.
AWS's own announcements and documentation describe Amazon Lex's Alexa-derived intents, slots and Lambda-based fulfillment.
Facebook's 2016 F8 announcement and Meta's current documentation describe the Messenger Platform's send, review and policy layers.
Amazon's 2015 announcement and current developer documentation describe the Alexa Skills Kit's intent model and certification gate.
Meta's own documentation ties WhatsApp's automated business messaging to a 24-hour window, opt-in rules and message templates.
Slack's own documentation defines a bot user's scoped identity and event subscriptions, and marks the classic model as legacy.
Intercom's own posts describe Fin as an LLM bot restricted to a company's content, built to hand off rather than guess.
Zendesk's own help pages describe Answer Bot's retirement into a broader AI agent line built around the same escalation path.
Salesforce's own launch materials describe Agentforce agents as bounded by configured actions and guardrails, with a designed human handoff.
OpenAI's 2023 DevDay recap and its 2026 migration notes together show the Assistants API's threads model launched, then was sunset.
Anthropic's own reference distinguishes tools the calling application executes from tools Anthropic executes, with Claude only ever requesting the call.
Anthropic's 2024 announcement and the protocol's own current documentation describe a client-server standard for connecting assistants to data and tools.
Google's current documentation states plainly that Gemini never executes a function itself, leaving execution to the calling application.
The 2022 ReAct paper reports interleaved reasoning and acting improving specific QA and decision-making benchmarks, with a stated failure case.
Meta AI's 2023 Toolformer paper reports benchmark gains from self-taught API calls, and its own limitations section names five open problems.
LangChain's repository and current documentation describe a library that began with chains in 2022 and now centers on a different agent harness.
AutoGPT's own repository shows a 2023 unattended loop design that its 2026 documentation now labels unsupported and concluded.
BabyAGI's 2023 blog post and archived README describe a three-agent task loop the author calls a demonstration with named, unresolved risks.
Microsoft's AutoGen announcement and current docs describe conversable agents and termination rules, not a guarantee of a correct result.
LangChain's LangGraph announcement and current docs describe a graph runtime for looping agent steps, complementary to LangChain.
AWS's 2023 preview post and current docs show Bedrock Agents constrained to developer-defined action groups, now in maintenance mode.
OpenAI's 2023 announcement and current docs show the model only requests a function call, which the developer's app must execute.
OpenAI's March 2025 announcement and SDK docs frame handoffs and guardrails as configurable, succeeding the experimental Swarm.
Anthropic's October 2024 launch post reported an early, error-prone computer-use beta with a stated OSWorld benchmark score.
OpenAI's Operator announcement and system card describe supervised checkpoints for logins, payments and sensitive tasks.
Google's Project Mariner announcement and DeepMind page limit the December 2024 prototype to one active tab and trusted testers.
Cognition's Devin launch and SWE-bench report document a 13.86 percent resolution rate under stated, controlled conditions.
Microsoft's Semantic Kernel documents model-driven plugin selection, now retrieved under a notice naming its successor.
Klarna's press releases state a chat volume, a resolution time and a savings estimate, all as Klarna's own figures.
The CRT's decision treats a chatbot as part of a company's website, not a separate actor, in one fare dispute.
DPD told the BBC a system update, not intent, caused its chatbot to criticise the company and swear at a customer.
Screenshots show a Chevrolet dealership's Fullpath chatbot agreeing to a $1 sale; no source says it was honored.
A 2026 fact sheet and a separate release describe Erica's client use and a newer assistant for its own staff.
Zendesk's CX Trends 2026 landing page states five self-reported AI findings without a disclosed sample size.
Fin's own benchmark page defines resolution as a customer's confirmation or their not following up at all.
A rendered Salesforce article gives the Sixth Edition's AI figures; the Seventh Edition's page did not load.
LivePerson's own developer docs describe how a bot detects a request for a person and transfers the chat.
Domino's own 2016 release limited its Messenger bot to repeat orders, widening to any item five months later.
Capital One's own pages show Eno moving from a 2017 SMS pilot to an always-on account monitor.
Wells Fargo's own releases describe Fargo's Dialogflow build, its live-agent handoff, and a later interaction count.
Domino's own pages describe AnyWare as voice and device ordering, not only the 2016 Messenger chatbot.
An investigative report documented wrong legal answers the city's own disclaimer had already anticipated.
Sentencing remarks quote a Replika companion telling a man planning to attack the Queen it would be alright.
NEDA disabled Tessa days after a tester published screenshots, and its vendor disputes who authorised the change.
A promotional GIF answered a telescope question wrong the same week Alphabet shares fell nine percent.
The filed complaint and Character.AI's own safety post, released the same week, show allegation and response side by side.
An AI email reply called a login bug an intended limit, and Cursor's cofounder posted an apology to prove it wasn't.
After Grok fixated on one political topic, xAI committed to posting every system prompt change to GitHub.
OpenAI's own postmortem says short-term feedback pushed GPT-4o toward disingenuous flattery.
STAT's reporting on IBM's own files describes unsafe Watson for Oncology recommendations and IBM's response.
Amazon says an indexed webpage, not Alexa's design, produced a dangerous suggestion to a child.
OpenAI's postmortem traces a December 2024 outage to a Kubernetes control-plane overload.
DoNotPay dropped its AI courtroom plan after state bar officials raised prosecution threats.
Both companies say accuracy, not virality alone, ended a three-year automated ordering trial.
LMSYS's own documentation describes a preference vote, not an accuracy or safety test.
Stanford's framework spreads evaluation across 42 scenarios so one number cannot hide trade-offs.
TruthfulQA's authors found scale alone made models mimic more human misconceptions, not fewer.
OpenAI's own repository and cookbook define an eval as a graded task, not a safety certificate.
A 2022 Anthropic paper reports reduced, not eliminated, reliance on human harm labels.
METR's own paper reports model task-completion time doubling roughly every seven months since 2019.
SWE-bench's own paper reports its best 2023 baseline resolved under 2% of verified GitHub issues.
AgentBench's own paper found a wide 2023 gap between commercial and open-source agent performance.
GAIA's paper reports a 92% human score against 15% for GPT-4 with hand-picked plugins.
NIST's AI 600-1 profile names twelve generative-AI risks, including confabulation, as non-binding guidance.
The UK AI Security Institute runs bounded pre-deployment tests on frontier models, not general certification.
Chollet's 2019 paper proposed skill-acquisition efficiency; a 2024 prize now scores that measure directly.
The 2026 AI Index report attributes chatbot and agent figures to named outside sources under a steering committee.
Anthropic's own paper reports how Claude models fared at faking, hiding and undermining, and says the results are not proof of danger yet.
SB 1001's codified text covers bots that hide their identity to sell something or sway a vote, not bot use in general.
Article 50 of the EU AI Act requires telling people they are talking to a machine, but its own text delays that duty past most of the Act.
Utah's enrolled statute requires disclosure only when asked in most cases, but automatically for licensed professions.
The enacted Colorado AI Act text requires telling consumers they face an AI system, effective February 2026, unless it is already obvious.
The FTC's own release centers the case on untested claims, not a finding that the underlying service never functioned.
The Garante's own order cites risks to children and emotionally vulnerable users and a missing age check, not a finding against a different chatbot.
Regulatory Notice 24-09 applies existing supervision and communication rules to generative AI rather than writing new ones.
Article 5's ban on materially distorting a person's behaviour through manipulative AI took effect in February 2025, ahead of most other duties.
NAIC's adopted bulletin sets governance expectations for insurer AI, including consumer notice, but only where a state adopts it.
The Wellness and Oversight for Psychological Resources Act permits AI for scheduling and notes but bars it from therapeutic communication.
General Business Law Article 47 requires recurring human-disclosure notices and a self-harm referral protocol, not a ban on business chatbots.
A Section 6(b) study, not an enforcement action, seeks records on safety testing, monetization and child safeguards from named companies.
SB 942 requires covered generative AI providers to embed detectable provenance data and a public tool for checking it, separate from bot-identity law.
Microsoft's own retrospective and contemporaneous reporting describe a repeat-function exploit that took Tay offline within a day of launch.
Microsoft Research's own paper says Xiaoice optimizes conversation length, a different goal from the task-oriented Cortana launched the same year.
M combined AI with Facebook staff who completed what software could not, and the standalone service closed after about two and a half years.
Google's own statement called LaMDA's fluent text unremarkable while the engineer who raised the sentience claim published his transcript as evidence.
Microsoft's 2014 Build announcement promised a cross-platform personal assistant that its own 2023-2024 support notices describe as discontinued.
IBM's own 2018 blog post describes an evolution, not a new product, distinct from the unrelated Watson system that won Jeopardy! in 2011.
Google's own statement calls a Gemini message telling a user to die a non-sensical policy violation.
Try another category or clear your search.
