For the third time in two months, a creative AI platform has made the same move: it routes your work to outside model vendors, makes you warrant your inputs comply with those vendors' rules, and disclaims responsibility for what they do. Runway in May, ComfyUI a week later, ElevenLabs this issue. The practical version: you are agreeing to route client material to global vendors you do not choose, cannot always identify, or hold responsible.
Vol 26, No 16 · June 18, 2026
The_Ledger
Eleven Labs Routes Around Responsibility, Artlist Removes Its Promise You Can Delete, A Coffee Date with AI
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.
ElevenLabs keeps discretion over which vendors run underneath Image and Video; the customer absorbs the compliance and liability for them.
ElevenLabs is known for voice. Image & Video is a newer, secondary product, and because the company does not build its own image or video models, it runs on outside vendors. That part is expected. What changed on June 11, 2026 is the terms. ElevenLabs now has you represent and warrant that your inputs comply with those vendors’ requirements, reserves the right to add, remove, or replace the vendors at any time, and disclaims liability for anything they do. The terms name none of them. They appear on a separate subprocessor page, and two are companies this newsletter already tracks: Runway and Kling.
In human terms: A post house runs a client’s unreleased footage through Image & Video to generate a few shots. Under the new terms, that footage may be processed by any of six outside vendors, in locations that include the United States, Singapore, the European Union, Indonesia, and Malaysia. The post house cannot tell the client which vendor handled the material, and ElevenLabs can change the vendor without notice. If that vendor suffers a security incident, ElevenLabs disclaims liability for it. And by submitting the footage, the post house has already promised ElevenLabs that the material complies with that vendor’s requirements, which the terms do not disclose. The practical bite lands at the next client security review: the studio can no longer answer a standard data-handling questionnaire cleanly, because it cannot name where the footage was processed, commit to a data-residency region, or confirm which vendors touched it. More below:
On June 16, 2026, Artlist updated its Privacy Policy and removed the sentence that told users they could delete their prompts, uploads, and outputs at any time. The access language stayed, and so did the carve-out that lets Artlist keep hashed copies of prompts. Only the deletion promise came out.
Prior: “Your text prompts, uploaded files, and the AI-generated outputs are accessible and can be deleted by you at any time via your “My Account” interface; please note that hashed versions of your text prompts may be retained for security and system integrity purposes.” Updated: “Your text prompts, uploaded files, and the AI-generated outputs are accessible via your “My Account” interface; please note that hashed versions of your text prompts may be retained for security and system integrity purposes.”
In human terms: A creator who uploaded reference images and generated outputs on Artlist last month, and expected to clear them out before handing a project to a client, will not find that commitment in the current Privacy Policy. The interface may still offer deletion, but the written promise that it does is gone.
Why this matters: A privacy policy is the commitment a platform will stand behind in writing. The prior version stated plainly that prompts, uploads, and outputs could be deleted on demand. The current version still says they are accessible and still allows Artlist to retain hashed copies of prompts, but it no longer states that you can delete them. We have not confirmed whether deletion still works in the account interface, so the accurate reading is narrow: the stated right was removed, not that deletion has been turned off. For creators who treat on-demand deletion as part of how they protect client confidentiality, the gap between a stated commitment and an unstated one is the part that matters. Standing alone, this is a small edit. In context, it continues a broader pattern in which platforms remove explicit user guarantees while leaving the product functionality unchanged.
OpenAI added two sentences to its Service Terms covering OpenAI Academy, its training program. If you sign in with an organization-linked account, OpenAI reserves the right to share your Academy activity with that organization’s administrators. A second sentence clarifies that business and enterprise terms do not govern OpenAI Academy, even when you sign in with a workspace account.
Added: “If you use an email address or account associated with an organization, including through Sign in with ChatGPT, OpenAI may share information about your OpenAI Academy activity, such as enrollment, engagement, progress, and completion status, with your organization’s owners or administrators.” Added: “For users with accounts associated with a business or enterprise workspace, the OpenAI Services Agreement or other business or enterprise terms do not apply to use of OpenAI Academy, even if a user uses Sign in with ChatGPT with their workspace account.”
Why this matters: If you take OpenAI Academy courses on a work account, your enrollment and completion data is visible to whoever administers that account. The second sentence is the less obvious point. People often assume their employer’s enterprise agreement, with its privacy and data protections, covers everything they do while signed in. For OpenAI Academy, it does not. The reach is narrow, limited to Academy and to organization-linked accounts, but the visibility is real.
Vol26, Issue 16 Smaller Changes
The Rest
Anthropic added documentation for a new model, Claude Opus 4.8, to its transparency hub, with a summary table and safety-evaluation summaries. No change to terms, privacy, or user rights.
A relative timestamp became a fixed date, May 14, 2026, and one entry in the related-links list was swapped. The moderation rules themselves are unchanged.
The “last updated” date was reconciled to June 1, 2026 to match the document footer. The EU Data Act switching and deletion provisions are unchanged.
A one-word grammatical edit, with no change in meaning.
The documentation and download links moved to developer.meta.com. The license terms are unchanged.
The company mailing address was updated.
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.
A Coffee Date With GenAI
Fitness Test
Prompt Tested
STRICT ACTION ORDER — MUST BE FOLLOWED EXACTLY. The man picks up the ceramic coffee mug nearest him and extends it toward the woman. The woman reaches forward and takes the mug directly from his hand. The mug transfer is fully completed before any interaction with the photograph occurs. After receiving the mug, she places it gently on the table beside her. Only after the mug has been received and set down does the woman pick up the photograph from the table. The woman extends the photograph toward the man. The photograph is a single physical photographic print with an image on only one side. The reverse side is plain white photographic paper with no image, text, markings, or graphics. The man reaches forward and accepts the photograph directly from her hand. The photograph transfer is fully completed before any further action occurs. The man studies the photograph. His eyes move naturally across details in the image. After examining the photograph, the man slowly raises his gaze and looks directly at the woman. A subtle emotional reaction passes across his face. Maintain accurate object permanence and continuity throughout. The coffee mug exchange must occur first. The photograph exchange must occur second. The photograph must never move before the mug transfer is completed. Natural eye contact, blinking, breathing, posture adjustments, facial micro-expressions, realistic finger placement, authentic hand-to-hand transfers, and physically correct object interaction. High temporal consistency. Premium cinematic realism. Subtle handheld camera movement. No object duplication, no object morphing, no skipped actions, and no changes to wardrobe, faces, environment, or table objects
Sequence of Events, Typography, Character Rendering
We are exposing multiple known challenges here: Strict sequence of events, cogent typography, character performance. Don't forget hands manipulating a floppy object with two distinct sides (photograph front/back).
Veo3.1
The model refused our camera direction, ignored the cup handoff to SR talent. The type on the coffee cup is generally consistent. Good motion in BG talent and street. Clearly the photo would need to be replaced by a VFX artist, which would probably be difficult, model did try to add a decent approximation of our start images photo content. Interestingly - this model added a wedding ring to SR talent, though our start image did not have this.
LTX2.3
Camera direction was followed. Cup typography not very good. The cup handoff did happen, sort of. Photo handoff happened, though awkwardly. Wedding ring added despite a lack of this in start image. BG action moves naturally (generally).
Seedance2.0
This really ignored our camera direction, but did an actually pretty decent job on the other aspects. BG car motion is not realistic and seems a vehicle pops into existence near end of clip. Photograph could be comp'ed by a human artist. By focusing on SL actor, we see the emotion is clearly computer-generated. Wedding ring appears again. Grading on a curve, this gets the green this week.
Wan2.2
No adherence to the camera instruction, photo moving without motivation, talent action is unrealistic and floaty. Cup handoff isn't that bad, though type is illegible. BG errors on back wall and SR female shirt texture is broken at times. Wedding ring not present - as instructed by our seed image.
Kling2.1pro
This wins the week for hilariously broken.
The Verdict
This week we have again not demonstrated a success story - fully synthetic, close to camera actors with complex interactions are just not possible reliably in a one-shot prompt like this. Perhaps with fine-tuning and multiple generations a model like Seedance might get there - but you'll still have the ultimate issue that synthetic actors aren't convincing nor very directable in a hero setup like this one. Plan to shoot this unless it's part of a quick montage where you don't stay on the shot for very long.