People may be ready to let an AI agent speak for them. They are much less comfortable when someone else’s agent speaks back.
A preprint submitted on August 18 reports this gap from two surveys of active users of a major dating platform. One surveyed 2,894 people about generative profile features; the other surveyed 2,617 people about autonomous conversational agents. The authors found that respondents’ average willingness to deploy their own agent was roughly three times higher than their willingness to engage with another person’s agent.
That is not a small interface problem. A service can have plenty of people willing to send automated messages and still fail if the people receiving them do not consent to the exchange.
Sending and receiving are different decisions
The researchers measured willingness to delegate communication and willingness to receive agent-mediated communication as related but distinct attitudes. Their model found a strong correlation between the two, yet it still fit the survey data better when sending and receiving were treated separately.
They then modeled what might happen if users were paired at random. Only 4% to 13% of directed pairs combined one person’s willingness to deploy an agent with the other person’s willingness to engage with it. Routing contacts toward people whose answers suggested higher receiver-side receptivity produced a 3.1-fold difference between the highest and lowest receptivity groups in an out-of-sample check.
Those figures are counterfactual estimates built from stated preferences. The study did not observe real agents arranging dates, measure relationship outcomes, or test whether people would behave as they answered. It is also a preprint, not peer-reviewed evidence.
The useful result is narrower and still important: permission from the sender does not settle permission for the receiver. Products that let agents initiate contact will need clear disclosure, receiver-side opt-in choices, and controls that do not treat automated outreach as neutral by default.
A self-improving agent may be learning the benchmark’s order
Another August 18 preprint re-examined two methods that let AI agents build a text memory from earlier tasks and use it on later ones. Researchers from Salesforce AI Research ran the methods repeatedly and shuffled the task order instead of relying on one default sequence.
The results changed substantially. Variance increased in 71% of the tested cases, and the difference between the best and worst runs reached as much as ten percentage points. More strikingly, the average result moved from a 1.5% improvement under the default order to a 4.5% decline after tasks were shuffled.
The authors argue that the default order acted like a hidden easy-to-hard curriculum. They also found memories containing plausible but unusable advice, such as recommending an API in a browser-only environment. Adding clearer rubrics and environment feedback recovered some performance, but significant gaps remained.
This does not show that agents cannot improve from experience. It shows that two published approaches were less stable under repeated and reordered tests than a single headline score might suggest. The paper released its code and data, making the result easier to inspect and challenge.
Meta wrote down promises that can be checked later
Meta published a long essay on August 10 describing its desired path to “personal superintelligence.” It says the company intends to offer capable personal agents broadly, provide free or affordable access, build a fully private mode that even Meta cannot access, and resume releasing some open-source models.
Axios reported two more governance commitments: an independent board would approve model-release safety criteria and review whether releases meet them, while paid access to extra computing capacity would use a dynamic auction mechanism.
The document is a company philosophy and product plan. It does not demonstrate that the promised private mode, future agents, access model, or social benefits exist today. Meta’s current products, data practices, release decisions, and board authority will provide the evidence for judging those promises later.
New South Wales asked for a review, not a final ban
On August 10, New South Wales Deputy Premier and Education Minister Prue Car asked the state’s education standards authority, NESA, to consider a moratorium on unsupervised take-home assessment while it reviews AI’s effect on learning and assessment.
The request matters because school-based assessment supplies half of a student’s Higher School Certificate result, and some of that work currently happens outside the classroom. The government asked NESA to consider a common approach for identifying inappropriate AI use and said possible interim changes could begin in Term 4 for the 2027 HSC class. Major art, design, and technology projects are expected to need exemptions or subject-specific treatment.
No blanket ban is final. The public release records a ministerial request and an active review. It also shows how the policy question is moving beyond whether software can detect AI-written text: assessment may need to be designed so teachers can authenticate the student’s own work.
Adobe put more than 70 tools behind one ChatGPT plugin
Adobe introduced a unified ChatGPT plugin on August 6 that it says draws on more than 70 tools across products including Photoshop, Firefly, Premiere, Acrobat, Lightroom, Illustrator, InDesign, Express, and Stock.
In Adobe’s described workflow, a user can ask for photo edits, campaign videos, design templates, social-media adaptations, or polished PDFs inside a conversation, then continue advanced work in an Adobe app. Adobe says the plugin is available globally on ChatGPT’s web and app surfaces. Its official demo video shows the conversational workflow.
The launch establishes availability and the connected tool list. Claims about higher quality, productivity, or a smoother creative process come from Adobe and still need independent use over time. A conversational front end can hide tool boundaries, but it does not remove the need to inspect source files, edits, licenses, and final output.
LinkedIn is reducing the reach of generic AI-like posts
One older platform update is worth keeping in view. LinkedIn said on June 4 that posts appearing AI-generated and lacking a clear perspective are less likely to be distributed widely beyond a person’s immediate network. It also said its systems act on automated comments posted at scale and replies that merely restate the original post.
LinkedIn reported that its initial test identified generic content correctly 94% of the time. That number is the company’s own evaluation; the notice does not publish an external audit, error breakdown, or appeal process.
The policy does not prohibit using AI to write. LinkedIn’s stated distinction is between assistance that still conveys a person’s perspective and generic or automated output with little added substance. For creators and professionals, the platform is making a human point of view part of distribution—not merely a style preference.
This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.