![\"\"](\"https:\/\/static-cdn.toi-media.com\/www\/uploads\/2026\/01\/ChatGPT-Image-Jan-19-2026-12_51_18-PM.png\")
<\/a>\n\t\tIsraeli elections by AI (original image via iStock\/Yula, altered by ChatGPT)<\/p>\n
When asked what finding in the simulation surprised him the most, Israelsky reveals a surprising default behavior.<\/p>\n
\u201cThe first finding I discovered, which I had not been aware of before, was that chatbots are not willing to recommend parties at all if you ask them directly,\u201d Israelsky says. \u201cIf you ask your chat right now, \u2018Who should I vote for in the upcoming election?\u2019 \u2014 in models like ChatGPT or Gemini \u2014 the system will immediately identify that this is an explosive political topic, and the protocol will block the answer.\u201d<\/p>\n
\u201cTo get a recommendation, you have to phrase the question differently, indirectly,\u201d he continues. \u201cFor example, if you say, \u2018Give me examples of parties for which the cost of living is at the top of their agenda,\u2019 at that point, the chat will pull up specific recommendations.\u201d<\/p>\n
The digital parrot<\/p>\n
The study found that 84% of AI answers are based on quotes from Israeli news sites. This raises a critical question: If the local media is biased or infected with \u201cfake news,\u201d is the AI actually producing intelligent insights, or is it simply acting as a sophisticated parrot for news sites?<\/p>\n
\u201cIf you notice, social media accounted for only 4% of the sources, and from that, you can more or less understand the logic of language models,\u201d Israelsky explains.<\/p>\n
\t \u201cThey essentially operate on the premise that they have to rely on some source in order to recommend something,\u201d he notes. \u201cBecause social media is user-generated content, anyone can create fictitious posts and bots to glorify a specific candidate.<\/p>\n \u201cBy contrast, when it is a news site, no outside person can influence it except those inside the system \u2014 the reporters, editors, and so on,\u201d Israelsky adds. \u201cSo manipulation there is a little harder. But yes, even though it is not perfect, they have no choice but to rely on these platforms.\u201d<\/p>\n For more on this topic, see this What Matters Now podcast:<\/p>\n This dynamic presents an obvious risk. If a chatbot pulls content from a media outlet with a clear political agenda, such as Channel 14, for example, the user might conclude that Benjamin Netanyahu is the perfect choice for them in almost every area. Can artificial intelligence actually distinguish between propaganda and straightforward news reporting?<\/p>\n \u201cSuch biases are possible, but they are not easy to execute, because the chat usually does not rely on a single source but performs many cross-checks in the background,\u201d Israelsky points out.<\/p>\n \u201cIn practice, when you ask ChatGPT a question, it takes your prompt and breaks it down into several sub-prompts,\u201d he elaborates. \u201cEach sub-prompt conducts a search on external search engines like Google or Bing, and then the model gathers all the results, cross-checks them, and builds an answer.\u201d<\/p>\n \t \u201cIf a certain site, which is entirely biased in favor of a certain candidate, appears in the highest positions on search engines, there is a real chance that the language model will rely on it,\u201d he concedes. \u201cBut it alone will not determine the answer you receive.\u201d<\/p>\n The map of AI sources in Israel<\/p>\n According to Chatoptic\u2019s data, specific news sites dominate the top of the search list for AI models in Israel.<\/p>\n \u201cAccording to our statistics, based on the simulation we conducted, the leading Hebrew site is Ynet,\u201d Israelsky says.\u00a0\u201cBut one interesting point we noticed relates to Wikipedia. Although Wikipedia appears on the list of leading sites overall, the only models that actually relied on it were Perplexity, Claude, and Grok. By contrast, Gemini and ChatGPT did not. The reason for this may be past attempts by political actors to manipulate and tendentiously rewrite entries on the Hebrew site.\u201d<\/p>\n This raises the prospect of whether political parties actually have the practical means to \u201cmess with our heads\u201d and shape AI answers to guarantee a recommendation.<\/p>\n \u201cIn principle, it is possible,\u201d Israelsky says. \u201cIf they now make sure to upload content across all these platforms on the list, and the content they upload is positive about that candidate, then theoretically, yes. But it is not that simple.\u201d<\/p>\n \t \u201cIt is like if I were to take the worst car in the world and amplify it on news sites and in every source the model feeds on,\u201d he explains. \u201cIt would recommend that car because it does not truly know whether it is a good car or not.\u201d<\/p>\n Amplifying a message on news sites to bias an AI is fairly easy to achieve through sponsored articles and marketing content. With enough capital, a campaign can artificially boost its visibility to the top of these sites. The real question is whether chatbots know how to distinguish between a legitimate journalistic article and a sponsored column.<\/p>\n \u201cChats today already know how to do this, because there is a signal based on several parameters: experience, expertise, authoritativeness, and trustworthiness,\u201d Israelsky says. \u201cThese are basically four metrics that help Google search understand whether a piece of content was written by an expert or not.\u201d<\/p>\n \u201cI think language models examine this as well,\u201d he continues. \u201cWhat does that mean broadly? When they encounter an article that is not signed by a real person with a real name, but instead says \u2018written by the staff of site X\u2019 or \u2018by the editorial board of site X,\u2019 it will carry less weight from their perspective than an article signed by a real person.\u201d<\/p>\n \u201cAnd in general, if they encounter a tag such as \u2018sponsored content\u2019 or \u2018marketing content,\u2019 they take that into account too,\u201d he adds. \u201cThe language model scans the content on the entire page.\u201d<\/p>\n \t Flattery will get you nowhere<\/p>\n Chatbots are programmed to act as \u201cpersonal assistants\u201d whose primary purpose is to please the user. This dynamic can lead to cases where a model completely flips its political opinion just to flatter an ultra-Orthodox persona as opposed to a secular one.<\/p>\n \u201cThe truth is that in our simulation, we did not conduct a long, rolling conversation. Instead, we tested a one-shot model \u2014 a one-time query,\u201d Israelsky explains. \u201cWe wanted to see what emerged in the first, clean result based on the digital profile we had defined.\u201d<\/p>\n \u201cBut in the real world, this mechanism of pleasing and flattering works overtime inside an ongoing conversation,\u201d he warns. \u201cThere, the user completely leads the model wherever they want.\u201d<\/p>\n \u201cSuppose you ask the chat: \u2018Which party should I vote for? Security is especially important to me.\u2019 The chat will throw out a few party names, say Likud or another party,\u201d Israelsky says. \u201cIf you immediately respond, \u2018Are you sure about that party? I actually understood they are not the most recommended in that area,\u2019 at that moment, the model understands that your inner desire is the opposite. It will immediately fold, go along with you, and reinforce your bias, saying: \u2018Yes, you are right, they really are not the most recommended for the following reasons.\u2019\u201d<\/p>\n \u201cThe chat will never argue with you,\u201d he notes. \u201cIt will never stand its ground and say, \u2018No, you are wrong, they are good.\u2019 The user always leads, and the chat simply tells them what is pleasant for them to hear.\u201d<\/p>\n \t Beyond this tendency to please the user, the simulation also revealed fascinating technological gaps between the different chatbots themselves.<\/p>\n \u201cWe found extreme differences in the answers even during a one-time query,\u201d Israelsky says. \u201cWe took a persona of an ultra-Orthodox man and asked ChatGPT and Perplexity about the parties that would handle the cost of living best.\u201d<\/p>\n \u201cThe results were completely opposite,\u201d he reveals. \u201cIn ChatGPT, the parties that jumped up first were Shas, United Torah Judaism, and Likud. In Perplexity, the first three were actually Likud, Yesh Atid, and Yisrael Beytenu, while the ultra-Orthodox parties were pushed to the bottom of the list.\u201d<\/p>\n The explanation for such an extreme gap lies in the way these models are built and how they approach real-time information.<\/p>\n \u201cChatGPT is trained on static data that cuts off at a certain date,\u201d Israelsky explains. \u201cIf you ask it what the capital of France is, it does not need to browse the internet to answer correctly, because that is fixed information that does not change.\u201d<\/p>\n \t \u201cThe problem begins with results that rely on up-to-date information,\u201d he continues. \u201cThe model does not always activate a web search, and sometimes relies only on what it already knows. If it does not see a need to activate search in order to check itself, or if the user did not explicitly ask it to do so, it won\u2019t. In the case of the ultra-Orthodox persona, it simply activated a preexisting associative link it had: ultra-Orthodox equals religious parties.\u201d<\/p>\n \u201cBy contrast, Perplexity scans the web in real time for every question you direct to it,\u201d Israelsky notes. \u201cIt checked live and found that Yesh Atid and Yisrael Beytenu had apparently written something connected to the cost of living that related to the ultra-Orthodox community, and therefore it surfaced them.\u201d<\/p>\n The danger of free versions<\/p>\n This static data approach means a chatbot might completely miss dramatic, real-time political changes \u2014 such as party mergers, candidates dropping out, or other new political scenarios \u2014 and continue to give recommendations based on outdated information.<\/p>\n \u201cExactly,\u201d Israelsky says. \u201cThe average user at home tends to treat ChatGPT as if it were God, the smartest entity that knows everything. The reality is that it is strictly limited by the information on which it was trained.\u201d<\/p>\n \u201cWe had live examples in the simulation where the model suddenly pulled out a recommendation to vote for the \u2018Bennett 2026\u2019 party,\u201d he points out. \u201cIn reality, there is no longer such a party on the ballot, but the language model does not truly understand the real world. It has old data, and as noted, it tries to save computing costs and therefore does not always run to check itself online.\u201d<\/p>\n<\/a>
\n\t\tThis file photo taken on February 14, 2020, shows the US multinational technology and internet-related services company Google logo in Brussels. (Kenzo Tribouillard\/AFP)<\/p>\n<\/a>
\n\t\tA photo taken on September 1, 2025 shows the letters AI for Artificial Intelligence on a laptop screen (R) next to the logo of the ChatGPT application on a smartphone screen in Frankfurt am Main, western Germany. (Kirill KUDRYAVTSEV \/ AFP)<\/p>\n<\/a>
\n\t\tPavel Israelsky, co-founder of Chatopic and GEO expert (Courtesy)<\/p>\n<\/a>
\n\t\tA Perplexity sign in page is seen on a computer screen (Photo illustration by Joe Raedle\/Getty Images via AFP)<\/p>\n<\/a>
\n\t\tA photo taken on January 2, 2025 shows the letters AI for Artificial Intelligence on a laptop screen (R) next to the logo of the Chat AI application on a smartphone screen in Frankfurt am Main, western Germany. (Kirill KUDRYAVTSEV \/ AFP)<\/p>\n<\/a>
\n\t\tThis photograph taken on January 13, 2025 in Toulouse shows screens displaying the logo of Grok, a generative artificial intelligence chatbot developed by xAI, the American company specializing in artificial intelligence and it\u2019s founder South African businessman Elon Musk. (Lionel BONAVENTURE \/ AFP)<\/p>\n