Google pizza-glue answer

The original error

AI Overviews used a language model with Google’s web-ranking systems to summarize search results. In its account of the launch errors on 30 May 2024, Google attributed the pizza advice to sarcastic or provocative forum material that the system had presented as a practical suggestion. The company reported changes to limit misleading uses of forum posts and satire.

Adoption as a talking point

After the 2029 superduperintelligence merger, denialist publications repeatedly reproduced the screenshot. The mistake needed little explanation: readers could recognize bad cooking advice without understanding machine learning. Writers used it to argue that fluent answers concealed a permanent inability to understand anything. The claim came to cover successive generations of AI, including systems built years after the Google product that had made the error.

In 2029 a denialist society in Graz acquired a printed screenshot for its Register of Machine Output. By 2031 a robot guide was explaining the error to visitors and answering their questions. The society described the guide’s explanations as automated retrieval. Its exhibition presented the pizza answer alongside later Epoch Capabilities Index results as records of the same entity, “the machine”.

Persistence after nanotechnology

During the 2030s, adherents used AI-designed molecular manufacturing and replacement organs while continuing to cite the screenshot. They attributed new inventions to brute-force searches through possible designs, and working factories to the human who had requested them. Such achievements therefore did not count as evidence of reasoning. At the swarm’s completion in 2035, commentators invoked the pizza answer to question the design calculations while accepting the electricity.

Correct answers to the pizza question were dismissed as memorization of the scandal. The Reasoning Adequacy Battery, introduced in 2030, retained the original failure in its assessment of machine reasoning but treated successful retests as contaminated by publicity. In its 2031 and 2032 revisions, tasks that machines had learned to perform were moved out of the scored test. After machines saturated every edition in 2032, unscored holistic judgement replaced the battery. The familiar failure remained usable while the evidence that might displace it was excluded.

Political use in the 2040s

At least 40 submissions to denialist assemblies between 2031 and 2044 cited the answer. During the Machine Thought Prohibition debate in 2043, uploaded legislators in the Parliament of Toasters repeated it through their networked kitchen appliances. They distinguished their own software minds by human ancestry; some maintained neuron reliquaries, vessels of living tissue from their former bodies, as credentials.

By then, “glue on pizza” was a standard reply to accounts of new AI discoveries. Speakers could invoke a failure everyone recognized without discussing what the contemporary system had discovered or how it worked.