Just now, ChatGPT rolled out the text watermark feature, and the running speed of GPT-6 has been increased by 50%.
Few AI vendors firmly believe that the products they release can "exchange quantity for quality", but OpenAI is exactly one of them.
Following its previous bold move of holding 12 consecutive days of conferences, yesterday, Thibault Sottiaux (commonly known as Tibo), head of OpenAI Codex, announced a 28-day commitment:
Every day going forward, the team will either release an update that delivers clear value to most Codex or ChatGPT Work users, or perform a full quota reset.
On Day 1 of the campaign, what OpenAI delivered is speed boost. For subscribed users using GPT-6 Astra and GPT-6.1 Sol, the default speed will be increased by around 50%, with no settings adjustments required from users.
On the same day, OpenAI also announced two changes in completely different directions. Part of the text outputs from ChatGPT and Codex will start to add invisible watermarks, and ChatGPT's advertising business is preparing to introduce a more intuitive image display format.
Taken together, these updates clearly illustrate OpenAI's current situation. It needs to make paid users feel that their subscriptions are more valuable, explain to regulators where AI-generated content comes from, and turn its growing user base into a business that advertisers are willing to pay for.
Apart from models, the daily affairs that AI platforms have to deal with are getting closer and closer to the familiar internet we know.
Daily updates, users expect usable improvements
There is a notable limit in Tibo's commitment: the improvements released every day need to be relevant to most Codex or Work users, and the effect must be significant enough.
In other words, a tiny feature hidden in a corner can hardly meet users' expectations for this commitment. For people who use AI to work every day, waiting time, operation convenience and sufficient quota will directly affect their willingness to continue paying.
The speed boost on Day 1 exactly responds to this demand.
According to Tibo's explanation, the default speed optimization for GPT-6 Astra and GPT-6.1 Sol covers OpenAI's subscription products, as well as partners that support Sign in with ChatGPT, including OpenCode, Pi, Amp and Devin.
Users who use subscription capabilities in these tools through their ChatGPT accounts are also covered by this adjustment. He said that users will gradually feel the changes within two hours after the announcement is released.
Later, Tibo gave a more specific figure: the generation speed has increased from 30 tokens per second to 50 tokens per second.
For work that requires continuous interaction with the model, speed boost still has practical significance. Writing code requires repeated file reading, modification, testing and error correction, and organizing materials may also go through multiple rounds of search and generation. After the single waiting time is shortened, the accumulated difference in experience will become more obvious.
However, compared with speed boost, quota reset may be more practical, and we also look forward to seeing more of them in these 28 days.
Leaving watermarks on text, and acknowledging the limitations of watermarks
Speed boost makes it easier for users to use AI, while text watermarks respond to another question: when generated content enters emails, articles, reports and public discussions, can the outside world still identify its connection with AI?
Also on the same day, OpenAI announced its solution to meet the EU's text provenance rules. According to its announcement, the EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way.
Global API customers can actively enable text watermarks for some models starting from the same day, and the default state remains off.
In the coming weeks, OpenAI will add invisible watermarks to eligible ChatGPT and Codex text outputs within the EU, covering all types of plans, but it will not set it as a global default function in the initial launch stage.
OpenAI calls this technology textGrain. It adds invisible statistical signals when the model selects words, and detection tools will look for these signals to judge whether a piece of text contains OpenAI watermarks. OpenAI also plans to open-source this technology.
The difficulty lies in that text is too easy to modify, and the expressions available for the model to choose from also vary with different content.
Under test conditions with a target false positive rate of 1%, for content such as psychology, the detection tool can identify about 80% of 200-token watermarked texts, and the recognition rate is about 95% when the length increases to 400 tokens. The expression space for mathematical content is smaller, so the detection performance is significantly lower.
Editing will also weaken the signal. In another set of 400-token text tests, when 10% of the words are replaced with synonyms, the detection rate will drop from about 92% to 66%; after the replacement proportion reaches 25%, the detection rate is only 17%.
These sets of figures show that text watermarks can provide clues, but there is still a long way to go to reliably judge the creation process of an article. OpenAI also said that in its benchmark tests for evaluating Astra, no meaningful performance difference was observed before and after adding watermarks, but these tests still cannot cover all actual writing scenarios.
The more critical boundary is that watermarks cannot measure human contribution.
A person who finishes the article by himself and then asks AI to adjust the expression, and a person who asks AI to generate the full text, may both involve the model processing the text. Detecting a watermark cannot explain how much judgment, editing and creativity the author has invested, nor can it be used to determine copyright, responsibility or the authenticity of the content.
Watermarks also will not be associated with specific users, accounts, prompts or conversations. Conversely, if no watermark is detected, it cannot prove that the text was completed independently by humans. Too short text, edited or translated text, or text from models that do not support watermarks may all make the signals unrecognizable.
Since false positives and false negatives still exist, OpenAI is currently only opening applications for detection tools to approved researchers and professional institutions, and has not fully opened it to the public at launch.
When ChatGPT reaches 1.2 billion people every week, ads start to show familiar forms
Compared with watermarks, visual ads make ordinary users feel the product changes more easily.
OpenAI said in the announcement on the same day that ChatGPT reaches 1.2 billion people every week. For a platform of this scale, commercial revenue beyond subscriptions has become part of product development.
The newly announced ad format will display product inspiration, usage methods or related experiences through images. The first batch of tests is planned to launch in the United States later this month, in cooperation with a group of initial advertisers, and the display scenarios are selected during the process of ChatGPT generating images.
According to OpenAI's description, ads will be clearly marked, kept separate from the images being generated, and ads will not affect the responses provided by ChatGPT.
Choosing image generation as the test scenario has its own commercial logic.
When users are conceiving room layout, event plans or lifestyles, images can make the usage scenarios of products more intuitive, and make it easier for brands to enter users' consideration range. Of course, whether it can bring a good enough experience still needs to wait for the test to give the answer.
OpenAI has also expanded its ad effectiveness measurement tools, cooperating with platforms such as Hightouch, Tealium and LiveRamp, to facilitate advertisers to send conversion data from their existing systems back to ChatGPT Ads, and introduce more attribution and effectiveness evaluation partners.
For advertisers, the real question that needs to be answered is whether users have made purchases after seeing the ads, and how many of these purchases would have happened anyway.
OpenAI announced several sets of early results provided by its partners.
DV Rockerbox said that WeightWatchers' attributed customer acquisition cost on ChatGPT Ads is 15.3% lower than its comprehensive paid search benchmark; WorkMagic said that 67% of the incremental purchases of health brand Dose came from brand new customers; Triple Whale said that 93% of the visitors that Portland Leather got from ChatGPT Ads were new visitors.
In addition, the position of ad display also requires new rules. OpenAI said it will evaluate whether conversations are suitable for displaying ads, avoid placing ads in emotionally vulnerable, sensitive or other inappropriate contexts, and explore brand suitability assessment with DoubleVerify and Integral Ad Science.
According to the announcement, these partners will evaluate relevant protection mechanisms in a controlled test environment and will not access the private conversations of real users.
Ads on traditional websites usually revolve around page content, while ChatGPT's ads appear next to a continuously changing conversation. The more the platform understands what users are trying to do at the moment, the more specific the business opportunities will be; similarly, users will also care more about what affects the suggestions they get.
OpenAI promises that ads will not affect responses, and we will wait and see what happens in the future.
The 28-day update commitment has just started, and what will be delivered every day next is still worth paying attention to. However, the several changes on the first day have made the scope of competition clearer. OpenAI needs to manage the efficiency of paid users, the transparency of generated content, and the commercial returns for advertisers at the same time.
The AI platform is becoming more and more mature, and it also increasingly needs to answer a familiar question: when it is both an assistant that helps you make decisions and a place where businesses compete for your attention, why do users continue to trust it?
References attached:
https://openai.com/index/eu-text-provenance/
https://openai.com/index/new-chatgpt-ads-format-and-measurement/
This article is from the WeChat official account "APPSO", written by the team that discovers tomorrow's products, and is published by 36Kr with authorization.