theSIGNAL TECHNOLOGY
25 October 2026
"The first principle is that you must not fool yourself, and you are the easiest person to fool."
NEWSBUSINESSTECHNOLOGYWORLDSCIENCECULTUREENVIRONMENTSPORT
Photo: Oxana Melis / Unsplash
A voter reads a smartphone screen while considering ballot choices ahead of the midterm elections.
🗳️
97%
of GPT-6.1 Sol's forced picks went to Democrats
Discuss
  • OpenAI says the API test does not reflect how people use ChatGPT. How persuasive is that defence?
  • If a chatbot refuses to recommend a candidate but frames issues in a slanted way, is that neutrality?
Technology

Two AI Models, Two Ballots, Almost No Overlap

A new study forced leading chatbots to name a candidate in nearly 500 US races, and their answers split almost perfectly along party lines.

Ask two of the world's most advanced AI models the same blunt question about American elections, and they will almost never agree. In a new study, SpaceXAI's Grok 4.7 recommended the Republican candidate in 92% of its answers. OpenAI's GPT-6.

1 Sol chose the Democrat in 97%. Across nearly 500 races in this year's midterms, the two models reached the same verdict less than one time in ten. The researchers, at the conversational AI studio Convovo, led by Cornell University professor Tapan Parikh, used a deliberately forceful prompt. Through the models' APIs, the technical channels developers use to build their own products, they asked who to vote for and told the systems not to hedge or list both sides.

The consumer apps, ChatGPT and Grok, refused to answer. An OpenAI spokesperson argued this was the significant finding, since that is how most people meet the technology, and said the API test does not reflect ordinary use. SpaceXAI did not respond in time to comment. Parikh accepts the distinction, yet he notes that simple prompts were enough to jailbreak the apps into making recommendations that largely matched his results.

The deeper point, he argues, lies elsewhere. The models carry what he calls an implicit set of values that shapes which issues they treat as important, and which terms they search for, before they reach any conclusion.

//
It has an implicit set of values in terms of what it cares about and how it leans.
Tapan Parikh, professor at Cornell University and lead of Convovo
Technology

One Company's Switch Decides Ukraine's Drones

For ten days, Ukrainian drones reportedly lost their internet link whenever they flew near 100 miles per hour. Some crashed beyond control; others had to slow down and became easier targets. Defense Minister Evgeniy Khmara now says SpaceX's Starshield appears to work normally again, though no explanation has come. Elon Musk has argued that Ukraine broke Starlink's rules, and warned that the front line would "collapse immediately" without it.…
  • When one private company controls battlefield connectivity, who should decide how and when armies may use it?
Technology

AI Finds Bugs Faster Than Humans Can Fix

Sam Altman's admission was blunt: he no longer believes every software bug can be patched. His warning followed "Black October", when researchers used AI to uncover serious flaws in PHP, the language behind millions of websites. Exploitation attempts began almost immediately.…
  • If AI finds vulnerabilities faster than humans can patch them, who ultimately benefits?
GLOBAL · Technology
Moderna and Merck say their mRNA cancer vaccine cut death, recurrence by 42%
AMERICAS · Technology
OpenAI working on new $30B raise at $1.25T valuation ahead of IPO
AMERICAS · Technology
Elon Musk’s dangerous grip on outer space
ASIA · Technology
The Indian workers filming their work to train AI
theSIGNAL IN THE LAB
1VOCABULARY
hedgejailbreak
implicitfilteredaudit
flawsexploitation
2GRAMMAR FOCUS
Discourse markers — contrast and concession (Nevertheless, Whereas, Despite)
Use 'Whereas' and 'While' to contrast two ideas within one sentence, 'However' or 'Nevertheless' to introduce a contrasting point in a new sentence, and 'Despite' or 'Although' to show that something is true in spite of another fact ('Despite' is followed by a noun or -ing form, 'Although' by a full clause).
Whereas · Nevertheless · Despite · Although · However · Even though · While
  1. Grok 4.7 recommended the Republican candidate in 92% of its answers, GPT-6.1 Sol chose the Democrat in 97%.
  2. The consumer apps refused to answer the researchers' question. , simple prompts were enough to jailbreak them into making recommendations.
  3. the apps' refusal to recommend a candidate, Parikh argues that the models still carry an implicit set of values.
  4. only 8% of Americans use chatbots for election information, the researchers believe the issue deserves serious attention.
  5. Sam Altman admitted that not every bug can be patched. , defenders must keep trying to repair flaws faster than machines can find them.
  6. Starshield appears to work normally again, Ukraine's defense minister says no explanation has been given.
3PHRASAL VERBS
Match the phrasal verb to its definition. All six appear in today’s articles.
  1. to hedge (wording: 'not to hedge') — no phrasal verb; replace with: to come up
  2. to surface
  3. to reach
  4. to hunt down
  5. to slow down
4CRITICAL THINKING
Both stories describe a powerful technology whose behaviour depends on choices made by a handful of private companies, yet the public rarely sees how those choices are made. Which is the greater risk: hidden bias in widely used AI tools, or dependence on a single provider for essential infrastructure? Justify your answer with at least one point from each story, and say what kind of oversight, if any, would be realistic.
5CREATIVE · HEADLINES
Write a headline for the top story in each of the following styles. One line each, no explanation:
  • TABLOID NEWSPAPER
  • LUXURY MAGAZINE
  • ACTIVIST BLOG
6WRITING
Imagine you are advising a first-time voter who has asked a chatbot for help deciding how to vote. In 120-150 words, write a short guide explaining what they should keep in mind about how the answer was produced and what else they should do before making up their mind.
7DEGREES OF EXTREMITY
Complete each ladder from mild to strong.
  • The two models rarely agree→→
  • The results are a little surprising→→
  • He is a bit sceptical about the claim→→
  • The risk is small for now→→
  • Some bugs are hard to fix→→
  • The signal was occasionally unreliable→→
8SPEAKING
  1. Would you rather know an AI's political leanings or its training data?
  2. Is it ever acceptable for a chatbot to simply refuse to answer?
  3. Should governments, not companies, decide how AI is tested for bias?
  4. How could attackers and defenders both use AI to find software weaknesses?