We are noticing similar changes are spreading through many company policies: responses to upcoming AI labeling laws, platforms building on their user base to add more services from 3rd party models - creating a morass of hurdles for users to navigate: who is responsible for what, where is my data going, and what IP is unprotected. We'll continue raising our hands to say these oft-changing policies are becoming more important as models become a permanent layer in our workflows.
Vol 26, No 19 · July 8, 2026
The_Ledger
Perplexity's Privacy Policy Drops Two Promises, We Test Character Consistency
The Notice
What changed in AI platform agreements this week
We are tracking over 60 companies and 280 policy documents daily - and growing. Things that govern how your data is handled, who owns what and who’s on the hook when something goes wrong. These are things you should be reading, but don't have the time for - plus they are super confusing.
If something truly concerns you, reach out to a lawyer - there are no guarantees we catch all changes, something vital to you may be missed or misinterpreted. Docs are diff'ed and summarized by multiple LLMs, then researched, edited and published by a non-lawyer human.
Perplexity swapped out its whole privacy policy this week. The old one, dated February 5, 2026, is gone. A new, shorter one dated July 2, 2026 took its place. It is about 40% shorter, and two things you could count on before are no longer written down: the setting that let you stop Perplexity from training its AI on your searches, and the promise that it would not train on your email. The new policy also covers Comet, Perplexity's web browser, for the first time.
In human terms: A user connected Gmail and Google Calendar to Perplexity's Comet browser so it could see their deadlines and pull up old email threads while they worked. One reason they felt okay doing that was a line in the old policy: Perplexity said it would not use your email to train its AI. This week they go back to check that the line is still there. It isn't. The setting that let them turn off AI training on their searches is gone from the policy too. Everything still works the same, and the browser still reads the inbox. But the written promises they relied on when they connected it are no longer in the document.
Why this matters: Those two lines were the reason a working professional could feel safe letting Comet into their inbox. Both are now missing from the text. Taking a promise out of a policy is not the same as changing what the product does behind the scenes, and we are careful about that below. Still, the direction is clear. A policy that used to spell out specific protections now uses vague wording and points you to a setting it no longer explains. If your job is to check what a tool actually promises before you use it, this tool now promises less than it did in June. Mechanics and details follow:
xAI reworked two documents on the same day: Grok's Terms of Service and its Acceptable Use Policy (the list of things you are not allowed to do). The Terms got shorter, and the long list of banned uses moved into the Acceptable Use Policy, which got longer. Along the way, xAI added rules for when Grok acts on its own, took a new license over any person's face or voice that shows up in your content, and widened the rights it holds over everything you upload. The banned uses did not vanish. They moved to the other document and grew.
In human terms: A editor uses Grok to help put together a pitch reel that includes footage of a client's on-camera actor. Under the new Terms, uploading that footage means xAI gets a license to use the actor's image, likeness, and voice, and it is now the editor’s job to have gotten every permission needed for that. Separately, a user tries Grok's new feature that browses websites and edits files on its own. When one of those automatic actions does something it shouldn't, the user finds out the Terms say xAI is not responsible for what those actions do. Neither the likeness rule nor the "not responsible" rule was in the version of the Terms the user first agreed to.
Why this matters: Two of these changes hit creators directly. First, the likeness rule means any face or voice in what you upload gets pulled into xAI's broader content license, and you are the one on the hook for getting permission. Second, the new "agent" feature lets Grok do things for you, including dealing with other companies and even banks, while xAI reserves the right to say it is not responsible for the results.
There is also a bright spot for our beat: the Acceptable Use Policy now bans removing the hidden "this is AI" tags and watermarks from content, which puts xAI in surprising company this week. Add it up, and xAI took more rights over your content and your likeness in one document while adding to your list of duties in the other. Details follow:
Magnific, (formerly Freepik), added a new kind of content to its Terms this week: "Magnific Originals," an AI-made streaming lineup of series, episodes, and trailers made with selected creators that you can only watch on the site. The new rules keep this content out of your normal license, limit it to adults, and, most interesting for our readers, ban using it to train AI and ban stripping out the tags that show where it came from.
In human terms: Someone browsing Magnific finds a polished AI-made short series and wonders if they can grab frames to use as training references for their own model. The new terms say no, for two reasons. The content is watch-only and not licensed for download, and using it or its metadata to train, fine-tune, or distill a model is banned outright. The same rule bans removing the "this is AI" label or the source tags that come with it.
Why this matters: An AI company is telling people they can't train AI on the AI videos it made itself, and can't strip out the tags that prove where those videos came from. That is an unusual stance, and it lines up with xAI's new tag-stripping ban to make this a two-company theme in a single week. Magnific also points to the EU's new AI law right in its everyday terms, and names itself as the one responsible for labeling the content as AI-made.
Some readers may find this all rather ironic. Details below:
ComfyUI added billing wording that can charge free users.
If a free user has a card on file "for identity verification, future upgrade purposes, or any other reason," the new clause says that user "expressly authorizes Comfy to charge that payment method for any usage that exceeds the applicable Free Tier limits," including overages from "technical factors."
It also says a heads-up warning "is not a condition of Comfy's right to charge for overages."
In plain terms: a card you added just to verify who you are can be charged if you go over the free limits, and they don't have to warn you first. For a tool with a big hobbyist and technical-artist following, that is worth knowing before you save a card.
Meta updated this policy to June 29, 2026, and two real changes sit under a pile of formatting noise.
First, the information used to personalize a child's Horizon experience got wider. The old text said "age, chosen language, and general location." The new text says "information such as age, chosen language, and general location, and people your child is connected with." So a child's contacts are now used to personalize the feed, and the words "information such as" leave the list open to grow.
Second, audio from Meta's AI Glasses can now go to the cloud. The old text said transcripts "are processed on your AI Glasses and in the Meta AI App." The new text says they "may be processed on your AI Glasses, in the Meta AI App or on Meta servers." Audio that used to stay on the device can now be sent to Meta's servers, and that is the part worth watching.
Vol26 Issue 19 Smaller Changes
The Rest
Artlist updated its Terms to July 2, 2026 and added a rule just for US users. The old version sent every dispute "uniquely to the Tel Aviv-Yaffo courts and no other court anywhere else." Now US users get New York law and New York courts. If you're in the US, the place you'd have to fight a dispute just moved a lot closer to home.
Kling added a new Section 4.9 that puts its official account in every user's Following list by default, framed as a handy way to catch "You can unfollow. It's a small on-by-default choice written into the terms, and it fits Kling's pattern of adding clauses without making noise about them.
Anthropic changed the models featured on its documentation page to Claude Fable 5 and Mythos 5, with a new note that Mythos 5 passes the company's internal safety limits in cybersecurity and biology, and that Fable 5 is the version released to the public behind automatic safety filters. It's not a terms change, but it's a useful signal for readers watching how AI labs describe which versions they hold back.
Zapier updated its compliance page to July 2, 2026 and moved its "Workspaces" feature from "Rolling out Q2 2026 ... Not yet generally available" to listed as live. A launch confirmation, no change to your rights.
The Exhibit
One prompt. Multiple platforms. Let's give notes.
Bringing decades of experience in feature film, episodic, and commercial production, we evaluate every output against what matters: does it meet the bar for professional use? Think of it as a dailies review for one-shot prompts.
Character Preservation
Fitness Test
Prompt Tested
-Extreme 85mm portrait of a 36-year-old woman standing in deep shadow beside a window. -Soft natural daylight enters from one side, gently illuminating only part of her face while the opposite side falls into rich cinematic shadow. -Detailed eyes and eyelashes, natural eye moisture, authentic facial anatomy, natural makeup, realistic skin translucency, subtle skin texture, individual hair strands, natural catchlights in the eyes. -Photorealistic. -Feature-film cinematography. -Ultra-high-end cinema photography. -Shallow depth of field. -Beautiful natural skin. -Emotional realism. -Soft roll-off between highlights and shadows. -Natural color grading. -Shot on a full-frame cinema camera with an 85mm lens at f/1.8.
Can Current Models Reproduce a Human Accurately
Our goal in this test is to evaluate which models have progressed to reliably recreate a targeted character. Older models are well known to struggle with character consistency and adherence to established likeness.
Seedance2.0
This model took a big liberty by blasting light onto our subject - without any reaction of the subjects pupils. Lips have a warble and overall the emotions read uncanny all the way (to me, the reviewer). That said, it's succeeded at preserving the character and does a great job with hair.
Production Readiness:
Veo3.1
Overall, this is the best of the group. Good job resolving into an OTS - and eyeline actually works pretty well. Skin detail is realistic and emotions are fairly naturalistic. Pupils are static in diameter, which may explain why they feel off a bit. That said, pretty good results.
Production Readiness:
Wan2.2
Good adherence to ingoing identity and environment. Facial performance is… odd - and not very comforting to the viewer, almost feels like a robot trying a few things out. Camera work is floaty.
Production Readiness:
LTX2.3
Though the facial texture is soft, there's something about the performance of this character I actually think is pretty good.
Production Readiness:
Kling
The model added an unrequested 'step back' by the avatar - but notice the resulting focus pull impacts vegetation edges - which is what I'd expect from a real camera. Facial details and emotion are pretty good, though identity softens a bit as she steps from camera.
Production Readiness:
The Verdict
This test demonstrates that prior concerns of genAI being able to faithfully reproduce the likeness of a supplied character are a thing of the past - all models passed this test. But, we now focus on likeness PERFORMANCE - not just the geometry or symmetry of a face. It's clear here that some models are better than others, and that eye details like pupils or direct-able micro expressions are the next barrier to cross.