Airtable - AI Labels

Views

Grid view

Hide fields

Filter

Group

Sort

Overlays

Filter tags

Caption

Byline

Title

Not labeled

Reveal

Interstitial

Annotation

Side-by-side

Comments

Pre-Roll

Username and bio

A visual label that covers the video or image
First Draft has guidance coming out on overlays

Deepfake-Tom-Warning

deepfake-overlay

Deepfake-label-full

Incentives for apps looking to promote
Noticeable as it is between the viewer and the media, rather than to the side
Cognitive science research on misinformation intervention recommends co-activating a false claim and a correction so a user doesn’t have to update memory later on (Thompson & Ecker 2018)
Overlays travel with the media, and (if done correctly) are difficult to remove

Only works if it is big enough, otherwise it might not be noticeable or readable, and it will be croppable
Not easily adaptable to audio deepfakes
More visual is covered with label - blocks media/makes it less accessible, especially if label is img and not text element

Who is applying the label? And are they the right person/entity?
Type and motivation of labels?
Is it hard-coded into the media or external?
How do add to audio deepfake? Add earmark/audio icons? But easier to trim out (interstitial) Can't do persistently without saying "this is a deepfake," could have underlying audio

Labels applied by platforms for the filters being used

AR-Spark-Gallery

Structured within the app ecosystem, citation

Lost as soon as someone saves asset and reshares on another platform (cf memecatfluffy example)

A textual label, such as a hashtag, that is to the side of the main media

image15.png

image23.png

Incentives for apps looking to promote
Noticeable as it is between the viewer and the media, rather than to the side

Only works if it is big enough, otherwise it might not be noticeable or readable, and it will be croppable

A mark of authorship by AI, eg via a byline on an article or a signature on a painting

Barrat-Obvious

Guardian-GPT-3

It’s a very clear attribution, so that AI’s involvement is explicit and follows existing conventions of authorship attribution rather than something specific

It can mischaracterize AI as the single author, originator, etc. - an ambiguous, excessive, deceptive attribution. ‘Attribution fallacy’ - “the propensity of people to uncritically accept implicit suggestions that some AI program or other is dealing with real-world situations.” Ekbia 2008, 8-9
See criticism of Guardian article
This can be, and often is, exploited to overstate the capability of AI, which is often the aim of its producers
The flipside is that it can conceal the role of humans

The title or headline identifies the media as AI-generated

Queen

Cat-Lawyer

How is this different from byline?

Not labeled, but use AI techniques to manipulate imagery

AI-Portrait-Lighting

Warriors-Art

Deep-Nostalgia-no-label

A notice that the media is AI-manipulated after it is consumed

Adele-Wardle

Obama-Peele

Boris-Fake

Nixon

Invokes a surprise, even shock, which can add emphasis

It could be missed, either not watched until the reveal, or shared before the reveal
It is croppable

A warning that interrupts, visually or aurally, the media part way through to notify users, possibly multiple times

Interstitial labels have been shown to be effective?
More noticeable
Reduces croppability

There are still gaps

cite examples of friction like Jigsaw if possible, Twitter prompts for articles

Interstitial in video, but also audio (in the way there are ads intermittently throughout podcasts, there could be “for those just tuning in, this is a deepfake”) - how often would you need that to be

Highlights artifacts of AI generated media

Bellingcat-Deepfake

Nixon-BehindScenes

Provides users skills to identify this media in the future / digital literacy

Requires forensic skills of user adding annotations
Takes time
Unstructured linking of annotations on media to notes

Original media and AI-manipulated media shown side by side

Busch-JLaw

Office-Phyllis

Nixon-SbS

Deep Nostalgia

Deep-Nostalgia-no-label

Convincing to users by allowing them to directly compare

Not always available, most applicable for face swapping type effects?
Requires extra editing and media size?
Can be fabricated? If convincing enough, could be denied — i.e. which (if any) is the original?

Even if the user doesn’t label the effect themselves, other users may start discussions about where a filter came from, effectively labeling the post and allowing users to look it up for themselves.

Memecatfluffy

deeptom-comment

Crowdsourced, allows users to look it up for themselves
Often the only means of addressing the ambiguity / uncertainty when not provided by the creator

No structured link to effect across platforms
Open to speculation, misattribution, deception, etc - no gatekeeping
The people labelling may not know the actual role of AI
Dependent on enough people encountering the media to comment

A notice that the media is AI-manipulated before it is consumed. This could be explicit verbal warning, or illustrated (eg by showing a real face and then the fake face)

Jones-Reckoning

iamjesserichards

Pre-activates the warning/notice, so that it is consumed with this understanding

Skippable, doesn’t work as well for long-form where people may start watching after the pre-roll
Croppable - the Alex Jones deepreckoning could easily be taken out of context

Handle or bio that indicates that all posts from account are AI manipulated

deeptomcruise-account

LilMiquela

wint-but-ai

Username is fairly visible
Bio is a space to add more detail, eg “wint but AI” is explained as ‘human-curated, AI-generated tweets using a GPT-2 model trained on dril.”

Bio not visible on posts, requires clicking through
A lot depends on the detail, sometimes description can be very subtle, as in the case of Miquela

13 records

Summary

Alert

Lorem ipsum

Okay