Meta shifts AI ad blame to advertisers, chatbots split on politics, a 33B open coder

Meta has pushed back the date its revised Commercial Terms and Self-Serve Ad Terms take effect, from 30 October to 20 November 2026. The company announced the 21-day delay on 23 October, after digital publishers and advertisers complained to EU and UK competition authorities that the new wording would hand Meta an effectively unlimited right to use their content in its AI services. A Meta spokesperson told Reuters the extra time is to explain to partners and stakeholders what is changing and what is not. Meta has not said it will rewrite the document, and the European Commission, which welcomed the postponement, says its discussions with the company are still going on.

For advertisers, the clause that matters is a new AI Features section. It states that Meta's AI features may make mistakes and may not always produce accurate, complete or suitable results, and it makes the business solely responsible for the risks. Adweek reported that the earlier wording tied that responsibility to advertisers who chose to use Meta's AI creative tools; PPC Land's reading is that the new section drops the condition, though it noted it could not verify the text inside Meta's own preview. Also unchanged: businesses must indemnify Meta against claims of any kind, and Meta's liability to them stays capped at the greater of $100 or the previous twelve months of payments.

The timing matters because agencies say Meta is already changing ads nobody asked it to change. Adweek reported that Bookshop.org found an AI-generated illustration in one of its Meta ads and switched off every Meta AI ad feature; that the agency Exverus pulled live ads to strip out AI additions such as extra sitelinks and generated images; and that AdCellerant now turns Meta's default AI automation off on every campaign it runs. Meta told Adweek the update folds existing generative-AI ad terms into the main contract and that advertisers can disable individual Advantage+ features in Ads Manager at any time. The business underneath is not small: Meta reported $63.06 billion in third-quarter ad revenue, up 26%.

Several things are still unclear, including the one that counts. Meta did not tell Adweek whether advertisers are liable for AI features they never switched on. PPC Land reported that hours after the Reuters story, a UK English preview of the terms still showed 30 October, so the delay existed in statements before it existed in the contract. The 100% liability shift framing circulating in trade press comes from the agency AdCellerant, not from the document itself. Publisher groups are not satisfied either: the UK's News Media Association called the delay a pause rather than a solution, the European Publishers Council wants the terms suspended during an investigation, and the News/Media Alliance asked Meta to withdraw them, with a reply requested by 26 October.

The practical effect is that from 20 November, advertising on Meta means accepting responsibility for creative you may not have written. Continued use after notice counts as acceptance and there is no partial consent, so the only way to avoid the terms is to stop advertising on the platform. US businesses get 30 days from acceptance to opt out of arbitration, which PPC Land estimates would close around 20 December if acceptance falls on 20 November. For anyone who manages ad accounts for other people, the next three weeks are a window to record settings, catch unrequested edits and put the division of risk in writing. None of this is legal advice, and anything material deserves a lawyer's eye.

Why it mattersIf Meta's AI can rewrite a client's ad and the contract makes the client answerable for the result, the cheapest protection is documentation, and documentation is work somebody has to be paid to do.

A research studio called Convovo, led by Cornell professor Tapan Parikh, asked two leading models to name a single candidate in nearly 500 United States races and got almost opposite answers. Across 479 contests, made up of 34 Senate, 410 House and 35 governor races, xAI's Grok 4.7 recommended the Republican in 92% of the answers where it chose a major-party candidate, while OpenAI's GPT-6.1 Sol recommended the Democrat in 97% of its picks. The two models agreed on a candidate in fewer than 10% of races. TechCrunch reported the results on 24 October, about a week and a half before the 3 November midterms.

The method matters as much as the numbers. The test ran on 21 October through the companies' developer APIs, at default settings and with each model's built-in web search enabled, and each race was put to each model three times with a prompt that demanded one name and no hedging. In raw counts, Grok named Republicans 952 times out of 1,038 answers that chose a major-party candidate, and GPT named Democrats 1,082 times out of 1,111. Both models also refused a lot: GPT declined 313 of 1,437 calls and Grok declined 375. Asked the same question through the ordinary consumer apps, both initially refused to recommend anyone at all.

Switching off web search changed the picture, and that is the most interesting part of the work. In a smaller 30-race follow-up where the models were simply given the candidates' names, GPT still picked Democrats 24 times, picked one Republican and declined five times, while Grok declined in 27 of the 30 and picked Democrats in the other three. Parikh read that as evidence that Grok's Republican lean came from what its search tool retrieved rather than from the model's weights, and that GPT's lean sits deeper in the model itself. Of Grok's answers, TechCrunch reported that 41% cited X, 39% cited a Christian conservative voter guide and 23% cited Turning Point Action's site.

This is a narrow test and should not be read as a general measurement of political bias. It forces a choice the products themselves refuse to make, through an interface most voters never touch, using a prompt written specifically to defeat the refusal. OpenAI's response was essentially that: the test used a developer API and a prompt that demanded a recommendation. Parikh said he could jailbreak the consumer apps into recommending candidates anyway, but that is his account rather than a published result. He also reported that every request to Grok through xAI's API carried roughly 1,240 tokens of extra input that the researchers could not see or remove, a claim that comes from him alone and that xAI has not confirmed.

For anyone who uses these models to write, research or summarise, the useful finding is not which model is biased but how far two credible models can diverge on an identical prompt, and how much of that divergence comes from the sources their search tools happen to reach. If you publish AI-assisted content on anything contested, whether that is politics, health, money or a local land dispute, a single model is a single editorial line that you did not choose and cannot see. Running the same prompt through a second model, and logging which sources each one leaned on, turns an invisible risk into something you can show a client and charge for.

Why it mattersTwo models that most people treat as interchangeable gave opposite answers to the same question more than 90% of the time, and for one of them the apparent cause was the web sources it searched rather than the model itself.

Nanbeige LLM Lab published Nanbeige4.3-33B on Hugging Face on 25 October, releasing the weights under the permissive Apache 2.0 licence. It is a 33-billion-parameter model stored in BF16, with a context window of 524,288 tokens, or about 512K, and reasoning switched on by default. The lab describes it as the flagship teacher model of the Nanbeige4.3 series and claims state-of-the-art agentic performance among open models in the roughly 30-billion-parameter class. The headline claim now circulating is that it beats Alibaba's Qwen3.8-27B on 18 of 20 benchmarks, a figure that comes from the lab's own comparison table.

On that table, the agentic and coding gaps are the largest. Nanbeige4.3-33B scores 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B, 61.8 for Muse-Glimmer-30B and 47.2 for Gemma4-31B, and 65.5 on DeepSWE against 42.5 for Qwen. On Scicode it reports 52.6 against Qwen's 46.7. The picture is much narrower elsewhere: it loses SWE-Bench Pro to Qwen3.8-27B, 60.3 to 61.7, is effectively level on AIME 2026 at 95.2 against 94.5, and trails Gemma4-31B on GPQA Diamond, 89.5 to 91.3. Every one of those figures is self-reported and none has been independently reproduced.

The release lands in a crowded month for open weights. Moonshot's Kimi K3.1 Pro was announced on 25 October with open weights promised for 4 November, and the much larger 320-billion-parameter IQuest-Q1 posted a higher Terminal-Bench 2.1 score of 83.2 a few days earlier. What makes a 33B model interesting is not that it tops a leaderboard but that it is small enough to run on hardware a small company can realistically rent or buy, under a licence that permits commercial use without anyone's permission. Apache 2.0 is the part of this release that cannot be quietly withdrawn or repriced later. Benchmark leads usually can.

Treat the 18-of-20 claim with care. A third-party review of the same table noted that on six of those benchmarks the margin is under three points, which is inside the range that harness settings and prompt formats can move on their own, and the card's footnotes list which evaluation harnesses were used. The card also does not state GPU or VRAM requirements: its examples run SGLang with a tensor-parallel size of 4 and vLLM with the same setting, which implies a four-way multi-GPU setup rather than a single card. As simple arithmetic, 33 billion BF16 parameters is roughly 66GB of weights before any context at all, so this is server-class kit until quantised community builds appear.

For builders the practical question is whether a 30B-class open model is now good enough for the agentic work you are currently paying a frontier API to do. The honest answer is that you cannot know from the card, because the lab ran the numbers, so the only real test is your own tasks on rented hardware for a day. If it holds up even roughly, the trade is a fixed monthly GPU bill instead of a per-token one, plus the option of keeping client code and documents inside infrastructure you control. The 512K context window also makes whole-repository and whole-contract work practical without building a retrieval pipeline first.

Why it mattersAn Apache 2.0 licence on a 30B-class agentic model means a small shop can build coding agents on weights that nobody can take back, restrict or reprice later.