Ethical Stack.

Gemini

Google assistant for conversation and research, with access to Search grounding and connected Google services.

Score calculation

(6 + 3 + 4 + 2 + 7 + 3) ÷ 6 = 4.2

Pressure target: −1 overall point. Overall score: 3.2/10.

6 of six dimensions rated. Unrated dimensions are excluded from the average.

Full scoring method ↗
Setup details

Gemini: cloud access. Free access; Google AI paid plans vary by country and date. Ratings describe the stated product, company or configuration in each finding.

What fits your needs

Consumer Gemini Apps; model, connected-app and activity features depend on account, region and settings.

Tasks

Chat, Google Search grounding and connected Google services can support research and work across text, images and files.

See the evidence

Access and billing

There is free access and paid Google AI plan access; limits and eligible models vary by country, account and plan.

See the evidence
Tradeoffs to expect

Route

Gemini Apps are cloud services that can use Google infrastructure and connected apps; consumer app use is distinct from running open models locally.

See the evidence

Activity and review

Keep Activity affects saving and use of activity; some human-reviewed chats and related data can remain for up to three years after deletion.

See the evidence
How to switch

Get started

Sign in at gemini.google.com or use the official app; inspect Gemini Apps Activity and connected-app permissions before sharing existing files.

See the evidence
Practical questions

Age and sensitive use

Google’s Generative AI policy prohibits CSAM, non-consensual intimate imagery, self-harm assistance, deceptive impersonation, and certain high-impact decisions without human supervision; account age eligibility depends on country.

See the evidence

Feedback and files

Users can mark a response Bad, add feedback, and optionally include the files/images from that conversation; general problems can be reported from Settings & help. Feedback may include prompts, responses and attached content.

See the evidence

Search and corrections

Report inaccurate or unsafe responses with Bad response or Report legal issue; feedback sends the associated conversation and any chosen attachments to Google for review.

See the evidence

Energy practices

6/10↑

Google published a May 2025 median Gemini Apps text-prompt estimate (0.24 Wh, 0.03 gCO₂e, 0.26 mL water) with methodology and limits. Google also reports company-wide annual renewable matching and offers Fast-Lite/model/thinking choices. These disclosures do not identify the current physical route of each prompt.

See the evidence

This is the strongest product-specific disclosure in the set: Google quantifies a historical median Gemini Apps text prompt and publishes method boundaries, including that estimates are point-in-time and not independent. Its annual 100% renewable matching is company portfolio scope and earns E1 under this rubric, not a per-prompt clean-power claim. Current model/thinking choices and Auto selection are user controls, but their actual physical savings are not quantified.

1 + 2 reporting + 1 electricity + 2 controls = 6/10. This grade covers documented practices and transparency; it does not measure the energy or emissions of a request.

Current prompt/model energy and water; request-region and serving facility; independently reviewed measurement; hourly clean-energy match; variation by prompt, modality, and plan; footprint of training and hardware.

Reporting2/3
Google published a dated Gemini Apps median text-prompt estimate for energy, carbon, and water with a comprehensive methodology. The figure is a company estimate for May 2025, not an independently reviewed current-query measurement.
Electricity sourcing1/3
Google reports annual renewable-electricity matching and regional carbon-free-energy progress. The reviewed consumer-product evidence does not allocate a specific region or site to the prompt, so it earns 1 point for corporate sourcing disclosure. Annual matching is separate from hourly local supply.
Resource controls2/3
Gemini users can choose Flash-Lite, Flash, Pro/Argon and reasoning depth, or enable Auto model selection that sends most prompts to Flash-Lite. Google reports greater usage for advanced models/higher thinking, but no measured prompt reduction from these controls.

Water disclosure

Google reports that its water stewardship projects replenished approximately 7.7 billion gallons in 2025, about 78% of its 2025 freshwater consumption. The company-wide figure is not a measurement of Gemini, Vertex AI, or a particular request.

Google-reported 2025 corporate operations and replenishment. The page does not allocate water consumption to a product, model, region-specific request, or inference workload.

Training data

3/10↑

Google model cards document model dependencies and training processing, while its broader data policy describes public, curated, proprietary and licensed sources. The app profile does not identify one fixed release or full training data.

See the evidence

There is a general sourcing and internal review policy, but current model cards largely defer training-dataset details to related releases and do not publish source manifests. No consumer prompt control is counted as a training-data rights safeguard.

1 + 1 source transparency + 1 permission + 0 creator controls = 3/10. This grades documented training-data safeguards, not a legal ruling.

Exact default model and its inherited base stages, named datasets, compatible permissions and source-specific creator remedies remain unresolved.

Source transparency1/3
Model cards show version lineage but do not identify the underlying named datasets.
Documented permission1/4
Google describes review of origins and licenses as a general process, without release-specific validation in these sources.
Creator controls0/2
No rights-holder exclusion or remedy tied to the assessed training sources is demonstrated.
Data-rights rubric ↗

Your data & privacy

4/10↑

Gemini Apps Activity may be stored by default for 18 months; users can change auto-delete and improvement settings. Human-reviewed data can be retained separately for up to three years.

See the evidence

The service offers user controls, but reviewer access and separate retention after deletion reduce the strength of the default privacy posture.

Openness

2/10↑

Gemini Apps is a hosted service with automatic or plan-dependent model selection; users may inspect the model used and some paid users can switch models. These controls do not make Gemini model weights or the serving stack available for local use.

See the evidence

The app offers model visibility and limited selection, but the reviewed product documentation doesn’t confirm downloadable Gemini weights, an open license, reproducibility, or self-hosting.

Worker protections

7/10↑

Google has a supplier-responsibility system with worker surveys, third-party audits and grievance access, alongside current worker reports describing pay, pace and support concerns among AI raters.

See the evidence

Google says its Supplier Code applies to supplier and staffing-partner work, extended workers can report concerns anonymously through Google’s helpline, and its 2026 FY2025 report describes supplier audits and worker surveys with follow-up. Separately, Alphabet Workers Union-CWA’s 2025 report and raters’ pay-parity letter describe low pay, benefits gaps, insecure contracts and inadequate task-time; The Guardian documented raters’ accounts of rapid task timing, distressing moderation work and limited mental-health support, while Google says suppliers employ these workers and set their employment terms. The company system merits credit, but findings from hardware/supplier programs do not prove every AI rater is covered or remedied.

1 + 2 pay & conditions + 2 worker voice & remedy + 2 evidence & follow-through = 7/10. This grades documented worker safeguards, not a measured rate of mistreatment.

The FY2025 supplier program is risk-prioritized and does not disclose whether its survey/audit cohort included GlobalLogic AI raters; Google’s own policy places wage/benefit setting with suppliers. Worker reports are specific but are not an adjudicated finding about every supplier or rater.

Pay & conditions2/3
Supplier standards and vetting apply to extended workforce and current AI-rater testimony documents concrete working conditions, but rates, coverage and implementation vary by supplier and some worker evidence is self-reported.
Worker voice & remedy2/3
Extended workers have access to Google’s anonymous helpline and supplier feedback channels; union organizing and direct worker reports document voice, though they also describe limited influence over conditions.
Evidence & follow-through2/3
Google reports third-party supplier audits, worker surveys, follow-ups and program improvements; these are company-reported and the published figures don’t confirm coverage of AI annotation contractors specifically.
Workers rubric ↗

Military ties

3/10Broad defense ties↑

Google has a current DoD AI/cloud contract and has demonstrated military training/exercise support, but the reviewed evidence doesn’t confirm combat targeting or strike-intelligence deployment.

See the evidence

Google Public Sector describes a $200 million ceiling DoD CDAO contract and access to AI/cloud capabilities, including classified environments. Its Mobility Guardian 2025 account describes an Air Force exercise/edge workflow processing air-defense data and generating insights for counter-UAS triggers; this is exercise and support evidence, not proof that Google AI selected targets or supplied combat intelligence in an actual operation. Google’s current AI Principles include human oversight and safety commitments but no categorical military-use prohibition. The broad DoD contract leaves substantial enablement potential, so classify as broad-defense and apply the cap; do not call the exercise direct combat use.

1 + 0 use restrictions + 1 enforcement & oversight + 2 relationship disclosure = 4, capped at 3/10. Known defense ties lower this rating; safeguards can improve it within the relationship’s cap.

Public sources do not disclose all task orders, operational use outcomes, customer-enforceable use restrictions, or independent oversight. The specific models and releases used in every defense workflow are not established.

Use restrictions0/4
The current AI Principles reviewed don’t confirm a binding weapons or harmful-surveillance prohibition for defense contracts; the older 2018 weaponization commitment is not current evidence.
Enforcement & oversight1/3
Google’s public AI Principles describe human oversight and safety practices but do not disclose military-specific independent review, enforcement, or remedies.
Relationship disclosure2/2
Google names the CDAO contract, its ceiling and the Mobility Guardian exercise workflow; the named public materials disclose no actual combat mission or current contract-level safeguards.

Google documents a broad DoD AI/cloud contract and military exercise support, but no confirmed operational use for target selection, strikes, or combat intelligence. Broad-defense cap applies. The relationship limits this grade to 3/10.

Military rubric ↗

BDS targets

Pressure target↑

Pressure target in the checked BNC campaign list.

See the evidence

The BNC campaign guide names Google / Alphabet, the company behind Gemini, as a pressure target.

Gemini: Image generation & editing

Cloud image generation and editing in Gemini, using Google’s Nano Banana image model family.

Score calculation

(6 + 3 + 3 + 2 + 7 + 3) ÷ 6 = 4.0

Pressure target: −1 overall point. Overall score: 3.0/10.

6 of six dimensions rated. Unrated dimensions are excluded from the average.

Full scoring method ↗

Energy practices

6/10↑

Google publishes company-wide data-center energy and clean-energy metrics and identifies Nano Banana 2.1’s model family. Gemini Apps offers model and output-resolution choices, but no per-image footprint.

See the evidence

Google publishes company-wide data-center energy and clean-energy metrics and identifies Nano Banana 2.1’s model family. Gemini Apps offers model and output-resolution choices, but no per-image footprint. The score describes documented measurement transparency and available controls, not environmental performance.

1 + 2 reporting + 1 electricity + 2 controls = 6/10. This grade covers documented practices and transparency; it does not measure the energy or emissions of a request.

Google’s environmental figures cover its company/data-center portfolio, not Nano Banana 2.1 or a particular image request. The reported global clean-energy share doesn’t confirm sourcing for the serving region or a specific route.

Reporting2/3
Google’s 2025 Environmental Report gives dated quantitative company data-center energy/emissions and clean-energy figures with methodology/endnotes. This is corporate scope, not image-service allocation, so D2 is the cap.
Electricity sourcing1/3
Google reports a 66% company-wide carbon-free energy share in 2024, measured on an hourly basis. This is a partial organization-wide claim without Nano Banana route/site allocation or independent verification, so it earns E1 only.
Resource controls2/3
Gemini Apps documents switching among Nano Banana Lite, Nano Banana 2, and Pro, and downloading at 1K or 2K resolution depending on plan. The model card also identifies bounded image output. These controls can limit model/output choice, but physical savings are unmeasured.

Water disclosure

Google reports that its water stewardship projects replenished approximately 7.7 billion gallons in 2025, about 78% of its 2025 freshwater consumption. The company-wide figure is not a measurement of Gemini, Vertex AI, or a particular request.

Google-reported 2025 corporate operations and replenishment. The page does not allocate water consumption to a product, model, region-specific request, or inference workload.

Training data

3/10↑

Google explicitly maps Gemini 3.1 Flash Image to Gemini 3 Flash and states the latter inherits Gemini 3 Pro. The cards defer dataset detail to predecessor cards, while Google’s general policy describes source review but not a release-level rights audit.

See the evidence

The inheritance chain is partly visible, but the model cards do not name the underlying datasets or licenses. The score applies only to this defined first-party image route and reflects broad sourcing policy, not claims about unverified Nano Banana variants.

1 + 1 source transparency + 1 permission + 0 creator controls = 3/10. This grades documented training-data safeguards, not a legal ruling.

The Gemini 3 Pro source manifest, exact consumer app mapping to this image release, and model-specific creator controls are not published in the reviewed evidence.

Source transparency1/3
Model cards make the image-to-base dependency explicit but defer the underlying named training sources.
Documented permission1/4
Google describes internal source/license review generally, but supplies no inspectable permission ledger for this inherited route.
Creator controls0/2
No rights-holder opt-out, removal or permission mechanism is shown for the specific inherited model training data.
Data-rights rubric ↗

Your data & privacy

3/10↑

For consumer Gemini Apps with Keep Activity on, prompts and shared images may be used to improve and train Google services, and some data is reviewed by service providers and retained up to three years. Users can turn Keep Activity off or use Temporary Chats, subject to separate service processing.

See the evidence

The consumer default permits model-improvement use and human review, while users have controls that reduce future review and training uses. The score applies to Gemini Apps consumer configuration, not Gemini Enterprise or API terms.

Openness

2/10↑

Nano Banana 2.1 is distributed through hosted Gemini, AI Studio, API, Search, Flow and other Google products; the current model card does not provide downloadable weights or a self-hosted implementation.

See the evidence

The model card improves disclosure about capabilities and distribution, but the image model remains proprietary and hosted. Google’s separate open Gemma models do not make Nano Banana 2.1 open.

Worker protections

7/10↑

Google has a supplier-responsibility system with worker surveys, third-party audits and grievance access, alongside current worker reports describing pay, pace and support concerns among AI raters.

See the evidence

Google says its Supplier Code applies to supplier and staffing-partner work, extended workers can report concerns anonymously through Google’s helpline, and its 2026 FY2025 report describes supplier audits and worker surveys with follow-up. Separately, Alphabet Workers Union-CWA’s 2025 report and raters’ pay-parity letter describe low pay, benefits gaps, insecure contracts and inadequate task-time; The Guardian documented raters’ accounts of rapid task timing, distressing moderation work and limited mental-health support, while Google says suppliers employ these workers and set their employment terms. The company system merits credit, but findings from hardware/supplier programs do not prove every AI rater is covered or remedied.

1 + 2 pay & conditions + 2 worker voice & remedy + 2 evidence & follow-through = 7/10. This grades documented worker safeguards, not a measured rate of mistreatment.

The FY2025 supplier program is risk-prioritized and does not disclose whether its survey/audit cohort included GlobalLogic AI raters; Google’s own policy places wage/benefit setting with suppliers. Worker reports are specific but are not an adjudicated finding about every supplier or rater.

Pay & conditions2/3
Supplier standards and vetting apply to extended workforce and current AI-rater testimony documents concrete working conditions, but rates, coverage and implementation vary by supplier and some worker evidence is self-reported.
Worker voice & remedy2/3
Extended workers have access to Google’s anonymous helpline and supplier feedback channels; union organizing and direct worker reports document voice, though they also describe limited influence over conditions.
Evidence & follow-through2/3
Google reports third-party supplier audits, worker surveys, follow-ups and program improvements; these are company-reported and the published figures don’t confirm coverage of AI annotation contractors specifically.
Workers rubric ↗

Military ties

3/10Broad defense ties↑

Google has a current DoD AI/cloud contract and has demonstrated military training/exercise support, but the reviewed evidence doesn’t confirm combat targeting or strike-intelligence deployment.

See the evidence

Google Public Sector describes a $200 million ceiling DoD CDAO contract and access to AI/cloud capabilities, including classified environments. Its Mobility Guardian 2025 account describes an Air Force exercise/edge workflow processing air-defense data and generating insights for counter-UAS triggers; this is exercise and support evidence, not proof that Google AI selected targets or supplied combat intelligence in an actual operation. Google’s current AI Principles include human oversight and safety commitments but no categorical military-use prohibition. The broad DoD contract leaves substantial enablement potential, so classify as broad-defense and apply the cap; do not call the exercise direct combat use.

1 + 0 use restrictions + 1 enforcement & oversight + 2 relationship disclosure = 4, capped at 3/10. Known defense ties lower this rating; safeguards can improve it within the relationship’s cap.

Public sources do not disclose all task orders, operational use outcomes, customer-enforceable use restrictions, or independent oversight. The specific models and releases used in every defense workflow are not established.

Use restrictions0/4
The current AI Principles reviewed don’t confirm a binding weapons or harmful-surveillance prohibition for defense contracts; the older 2018 weaponization commitment is not current evidence.
Enforcement & oversight1/3
Google’s public AI Principles describe human oversight and safety practices but do not disclose military-specific independent review, enforcement, or remedies.
Relationship disclosure2/2
Google names the CDAO contract, its ceiling and the Mobility Guardian exercise workflow; the named public materials disclose no actual combat mission or current contract-level safeguards.

Google documents a broad DoD AI/cloud contract and military exercise support, but no confirmed operational use for target selection, strikes, or combat intelligence. Broad-defense cap applies. The relationship limits this grade to 3/10.

Military rubric ↗

BDS targets

Pressure target↑

Google is named as a pressure target in the BNC campaign sources.

See the evidence

Google is named by the BNC as a pressure target. BDS campaign sources state their own position; this classification does not assess the product or independently adjudicate the underlying claims.

What fits your needs

Gemini Apps image generation and editing context; model access, activity controls and limits depend on account, region and plan.

Tasks

Generate and edit images in Gemini Apps using text or image prompts; output and feature access depend on current model availability.

See the evidence

Provenance

Google's SynthID can mark Google-generated media and Gemini can check supported content credentials, with stated detection limits.

See the evidence
Tradeoffs to expect

Route

Consumer generation runs through Google-hosted Gemini Apps; local models and developer APIs are separate routes with different terms.

See the evidence
How to switch

Get started

Open Gemini Apps, select image creation/editing, and check plan/model eligibility and Keep Activity before sharing uploaded references.

See the evidence
Practical questions

Age and image boundary

Gemini Apps age eligibility depends on country/account. Google policy prohibits NCII, CSAM and use of personal data or biometrics without legally required consent.

See the evidence

Likeness and sexualized content

Google policy prohibits impersonation intended to deceive, explicit sexual content, NCII and privacy-rights violations; these policy rules don’t confirm consent for a particular uploaded likeness.

See the evidence

Labels and verification

SynthID verifies Google AI media only; Content Credentials require supported metadata and indicate provenance history, not authenticity or rights clearance.

See the evidence

Commercial use

Google’s generative AI policy applies to image generation; it does not provide a blanket copyright clearance or output warranty.

See the evidence

More from Google

All sources

Side by side