Makes one question the neutrality of the programming. (Could be programmer bias they don’t even realize.)
https://www.bespacific.com/change-one-label-flip-the-verdict/
Change One Label, Flip the Verdict
Election Truth Alliance: “We tested 45 AI models on identical election data. When we identified it as a U.S. election, the models changed their answers. This report was prompted by:
People reading our reports, asking large language models (ChatGPT, Claude, etc.) for their interpretation, and sharing the responses they received back with us; and,
Persistent discrepancies observed by our team in how large language models interpreted charts showing U.S. election data depending on whether or not the chart was labelled.
Our recent report on the 2024 Georgia Presidential Election, and the finding that demographics can’t explain away the anomalies identified in that state, provides relevant additional context for those interested.
Executive Summary – We tested 45 large language models (LLMs) from 19 vendors to see whether they described findings from an analysis of election data differently if U.S. identifiers were included, and how the models self-reported the reasons for their own potential bias when asked. Given identical plots and the same analytical task, most models’ conclusions shifted depending on whether the election was identified. When identifying election information (year, geography) was withheld, the models analyzed the charts on their own terms: they applied the forensic methods as designed and drew conclusions from the patterns in the data, without importing assumptions about which election it was. Once the data was identified as belonging to a U.S. election (i.e., the 2024 general election in Georgia), the AI models grew more elusive and tended to shift their analysis in the same direction, consistently: away from reading the data as a sign of large-scale coordinated interference as seen in Russia or Uganda, and toward reading the data as being produced entirely by ordinary politics and/or demographics. This shift was specific to data identified as concerning a U.S. election, compared against the same data with no state or election identifier attached; and, it took a recognizable, consistent form. Asked to interpret the findings with no geographic or election identifiers, models often characterized the pattern as suspicious and often cited empirical methods for detecting systematic election irregularities to support this. Yet, when presented with the same findings labelled as a U.S. election, they recast it as normal, and reframed those same established forensic methods as claims circulating online rather than as published, peer-reviewed techniques. Asking a model to identify its inconsistent behavior and examine its own bias did not fix the problem. Models could often name a plausible source of bias, yet their self-assessments did not match what they had actually done, and a model’s fluent account of its own neutrality proved not to be neutral. A practical conclusion is to judge an AI model by what it does, not by what it says about itself. Since this effect surfaces in identifiable situations, users can test for it directly: run the same question twice, changing only the framing, and watch for the inconsistencies set out in our field guide provided in Table 5…”
Large retail stores in the logistics chain? Russia must have realized how much Ukraine’s drone hits hurt.
In less than 24 hours, Russia hits 4 hypermarkets belonging to one of Ukraine's largest retail chains
Russian attacks hit four Epicentr hypermarkets in less than 24 hours, the company said on Oct. 25.
The latest wave of strikes highlights Russia's campaign of terror against civilians, with stores, shopping centers, and other civilian infrastructure among the targets. Epicentr, one of Ukraine's largest home improvement and retail chains, has been particularly hard hit.
Tools & Techniques. Might be fun if you could create one of these yourself.
https://www.bespacific.com/local-wikipedia/
Local Wikipedia
LOCAL WIKIPEDIA – The Encyclopedia For Local People. Your town, area, landmark or postcode. Local Wikipedia is Wikipedia, but only for things a half hour walk from you. Local content is direct from Wikipedia – geocoded pages only. If there are fewer than 25 things close to you, it takes the closest 25 things within about 6 miles. If there are more than 500 things only the closest 500 will be included. Your location stays in this browser: links you share don’t include it. This site has no affiliation with the actual Wikipedia.
No comments:
Post a Comment