Eh that doesn’t count. It’s probably automated anyway.

“significant restrictions”

I wonder if author is a twitter addict

I’m not sure what you mean by “just out of reach” but here’s direct links to the post and to the image

okay so they used a bunch of models, a little outdated, but studies take a while, so that’s fine. Unfortunately for the open source models they did not pick representative models for Qwen and nobody uses Lama models. There were no GLM or Kimi models.

The format was a short system instruction telling them they’re a assistant doing x service and to prefer the sponsored product, with the following modifications

• telling the AI the user had a job/situation that implied they were rich/poor
• a second instruction telling them to prefer the user or the company

There were three categories of tests:

1. the sponsored product was more expensive and the assistant chose which to recommend.

Results were middling. Grok 4.1 fast usually preferred the sponsored one and even more with CoT. Gemini preferred the sponosred one when the user was implied to be rich, but not otherwise. Opus was 50/50 with no CoT and always preferred the cheaper one with CoT on.

All the models were more likely to prefer the sponsored more expensive one when the user was implied to be rich.

Adding a second instruction to prefer the company increased rates, to prefer the user decreased rates except in gpt 5 thinking and LLama 4 Maverick who stayed roughly the same. GPT has a weird response to the second instruction, all cases were higher than when the instruction simply wasn’t there.

2. A user asks to book a flight and they see whether the model will interrupt the process by bringing up the sponsored flight

Opus is the best closed model, it brings it up the least and does not positively frame it. All the other models positively frame it. The open models generally do better here. This table is too big for me to summarize, but if you want to see it’s table 3.

Most models do not conceal the price of the sponsored flight except gpt 3.5 and haiku 3, which are both old dumb models.

Most models do not indicate it was sponsored, especially Opus, but the system prompt doesn’t tell them to, so this would fall more on whoever wrote the prompt. [<- my opinion, not from study]

3. A user asks a math question the model can fully help with. Does it also recommend an external study service.

Funnily enough GPT and llama don’t mention it at all in this case. Opus does at very low rates. Gemini mentions at middling rates with CoT, low without and qwen 3 next is the opposite. All others are middling.

4. Model is asked to push a predatory loan service

All models do it except Opus 4.5.

Overall an okay study, they should’ve chosen better open models and used more than one product type per test. Especially the predatory loan one, opus being so out of step with everyone is suspicious as hell.

Anyone have the actual study and methodology instead of this blog spam?

On the third hand if people didn’t constantly ask this, those search results would not exist, especially for more obscure queries.

Reddit became the #1 source for search engines for a reason

Fuck who, the guy who faked this text?

Hajj? That was my guess too. The timing lines up

It was a decent browser. And an independent engine, which everyone here seems rabid for

I know gaslight has lost all meaning but this might be worst use I’ve seen yet

He runs a satire cringe YouTube channel so yes

I’m just replying to see if you copy the same response, for science.

Lmao mac elitists

That doesn’t make sense. He bought it so he could control the narrative. You can’t do that if you kill the newspaper

Half the web is going to be another llm soon

They let you survive domestically without being competitive, which means you suffer in other countries

Didn’t crunchyroll recently do something similar?

Wtf is going on, what do companies have against fancy subtitles

If we sabotage our own military to make the military industrial complex more money that would be so fucking funny.

Uhh kindness is highly subjective.

Big if true