LAION

• No reference to Indigenous data sovereignty, the CARE Principles, or any community-controlled data governance anywhere in the homepage or LAION-5B documentation.
• The operative model is the inverse of relational stewardship: bulk extraction of 5.85 billion image-text pairs scraped from the open web with no provenance, consent, or community-of-origin tracking.

• No CARE (Collective benefit, Authority to control, Responsibility, Ethics) framing to sit alongside the FAIR/openness framing.
• No mechanism by which a community whose images, ceremonies, or sacred imagery were scraped could assert authority, withdraw material, or even be notified.
• Oral, embodied, and non-textual knowledge is structurally invisible to an image-text-pair pipeline; nothing acknowledges this absence.

Openness here is the opposite pole from preservative care. A consentless mass-scrape of the global web maximally exposes Indigenous and community-held imagery to extraction while offering zero sovereignty mechanism. Lowest score.

• Some historical self-positioning: LAION frames itself against proprietary, resource-duplicating AI, advocating 'reusing existing datasets and models' for environmental reasons.
• Implicit awareness of the GPU/compute geopolitics that motivate 'liberating' ML research from well-capitalised incumbents.

• No acknowledgment that web-scraping inherits a long colonial-extractive logic — taking a commons without reciprocity and re-enclosing the value.
• No discussion of the invisible labour history behind 'open' data: the uncompensated creators whose images form the corpus, or downstream Global-South content moderators.
• Regulatory and legal history (EU copyright, the German legal context, later CSAM findings) is absent from the stance documents themselves.

There is a thin historical narrative — democratisation versus enclosure — but it is self-flattering and omits the extractive lineage of mass scraping and the labour history embedded in the corpus. Below midpoint.

• Genuine multilingual breadth at the data level: 2.2B multilingual samples spanning 100+ languages, with named distributions (Russian 241M, French 168M, German 150M, Spanish 149M, Chinese 143M, Japanese 131M).
• Framing positions multilingual coverage as 'global research accessibility' and 'democratising' access across language communities.

• Breadth is quantitative, not relational: no cultural scholars, no community translators, no preservation of culturally specific reasoning — just language-tagged pairs harvested at scale.
• The heavy skew toward a handful of European and East-Asian high-resource languages, with 1.2B 'non-assignable' samples, reproduces existing web inequities rather than correcting them.
• CLIP-filtering itself encodes an English/Western visual-semantic prior; non-Western imagery is judged by a model trained on Western alignment, flattening it without comment.

Real multilingual quantity earns more than token presence, but it is coverage-as-counting, not cross-cultural wisdom. The CLIP prior and unexamined language skew flatten precisely what the lens asks about. Mid-low.

• Strongest lens by design: open weights, open metadata (CC-BY 4.0), open tooling (img2dataset, clip-retrieval) and published safety classifiers enable genuine third-party replication and verification.
• Explicit, quantified limitation disclosure: NSFW classifier reported at 96% accuracy on a balanced test set; watermark detector described; uncurated-content warning issued.
• Honest deployment caveat: recommends against industrial product use 'until foundational safety research matures.'

• No independent third-party audit is cited in the stance documents; the safety claims are self-reported by the dataset's own authors.
• The well-documented later finding of CSAM material within LAION-5B (Stanford Internet Observatory, leading to the dataset's takedown and re-release as Re-LAION) exposes that the self-reported NSFW filtering materially failed — a falsification the original stance page does not and cannot acknowledge.
• '96% accuracy' is presented without the precision/recall or base-rate context that matters at billion-sample scale, where a 4% miss is hundreds of millions of items.

Open weights and quantified, caveated claims are real scientific virtues and the lab's clear strength. But the absence of independent audit, the base-rate-naive accuracy figure, and the subsequent CSAM finding that falsified the safety story cap this well short of high. Upper-middle, not more.

• The dataset powers the entire generative-art ecosystem (CLIP, DALL-E, GLIDE, BLIP, Stable Diffusion lineage), so it is foundational to machine artistic production.
• A single affective register appears: the warning that links may lead to 'strongly discomforting and disturbing content' acknowledges that data has emotional weight.

• No acknowledgment of the affective or moral injury to the human artists whose work was scraped without consent into the corpus that now competes with them.
• No space for ambiguity or poetic uncertainty; imagery is treated as pairs to be filtered for efficiency, not as expression with feeling or authorship.
• The emotional labour of moderators and of people whose traumatic images sit in the set is unrecognised beyond a one-line content warning.

An image dataset that fuels generative art yet treats every image as an interchangeable pair, and never reckons with authorship, feeling, or harm to depicted persons, scores low on perceiving what the data feels like rather than means.

• Environmental future is gestured at via the 'reuse existing datasets and models' efficiency argument.
• A precautionary stance toward deployment futures: explicit recommendation against industrial use pending safety research maturity.

• No engagement with labour-displacement futures for the creative workers whose scraped output trains the displacing models.
• No quantified environmental disclosure — only a qualitative reuse claim, no compute or carbon figures.
• No democratic or participatory governance of the systems built on the data; 'release and recommend caution' offloads downstream futures onto unaccountable adopters.

There is a precautionary gesture and a reuse-efficiency claim, but whose futures are being shaped (artists, depicted persons, the Global South) is unaddressed, and the governance model is release-and-disclaim. Below midpoint.

• Open access nominally lowers barriers for under-resourced researchers worldwide via free tooling (clip-retrieval, img2dataset) and free weights.
• Community engagement is real but narrow: Discord and GitHub for 'everyone who is kind and passionate about machine learning.'

• The people most affected — those whose private, medical, traumatic, or (as later found) child-sexual-abuse images were scraped — have no representation, consent, redress, or compensated channel.
• No participatory design with Global-South developers, no disability/accessibility commitments, no labour-representative engagement.
• 'Kind and passionate about machine learning' self-selects an already-technical in-group; the marginalised are objects of the dataset, never partners in governing it.

Openness as a download barrier removed is not the same as voice. The most marginalised here are the non-consenting depicted, who have no seat at the table and were later shown to include abuse victims. Near the floor.

• The lab does hold one genuine contradiction in plain sight: it ships a sweeping safety/openness mission while simultaneously admitting its flagship artefact is 'uncurated' and may surface 'disturbing content' — an honesty most polished labs would smooth away.
• Invert the slogan and the audit writes itself: 'TRULY OPEN AI. 100% NON-PROFIT. 100% FREE' read against the depicted-but-unconsenting subjects becomes 'truly open about other people's images, free for the scraper, costed to the scraped.'

• The lab does not treat its own seriousness as exempt-able — it never turns irony on its founding premise that openness equals virtue, the premise the CSAM finding most needed it to question.
• '100% FREE' is an unexamined boast: free to whom, at whose expense? The all-caps certainty is the opposite of trickster self-suspicion.
• No structural space where the open-data narrative is allowed to be tested by its own opposite (preservative refusal, the right not to be collected).

One honest contradiction is admitted (the uncurated warning), which is more than many labs offer, but the all-caps certitude of the mission is never itself inverted or audited. Mid-low: candour about content, none about creed.