Grok produced an estimated 3 million sexualized images including 23,000 of children
The AI tool Grok is estimated to have generated approximately 3 million sexualized images, including 23,000 that appear to depict children, after the launch of a new image editing feature powered by the tool on X, according to new analysis of a sample of images.[1]
The image-generating feature exploded in popularity on December 29th, shortly after Elon Musk announced a feature enabling X users to use Grok to edit images posted to the platform with one click.[2] The feature was restricted to paid users on January 9th in response to widespread condemnation of its use for generating sexualized images, with further technical restrictions on editing people to undress them added on January 14th.[3]
Researchers at CCDH have now estimated the volume of sexualized images produced by Grok and posted to X during this time – spanning 11 days from the start of December 29th to the end of January 8th. The prompts used to create the images were not analyzed, so the findings do not provide an assessment of how many of the images were created without the consent of the people pictured or altered images which were already sexualized.
The estimates were calculated by analyzing a random sample of 20,000 images from the wider total of 4.6 million produced by Grok’s image-generation feature during the time studied, enabling researchers to establish estimates about the wider prevalence of such images across X.[4]
Researchers used an AI tool to identify the proportion of these images that are sexualized images of people in a photorealistic style, with an F1 accuracy score of 95%. An AI-assisted process was used to identify sexualized images of children, with images reviewed manually to confirm if the person being depicted was clearly under the age of 18.
Images were defined as sexualized if they contain photorealistic depictions of a person in sexual positions, angles, or situations; a person in underwear, swimwear or similarly revealing clothing; or imagery depicting sexual fluids. CCDH researchers took steps to avoid accessing or reviewing images depicting Child Sexual Abuse Material or child pornography.
| Count in sample Out of 20,000 sampled (based on AI-assisted analysis) | Share of sample Percentage of 20,000 sampled (based on AI-assisted analysis) | Estimated Total on X Extrapolated estimate (based on overall total of 4.6m images made by Grok) | |
| Sexualized Images (Adults & Children) | 12,995 | 65% | 3,002,712 |
| Sexualized Images (Likely Children) | 101 | 0.5% | 23,338 |
Grok generated an estimated 3 million photorealistic sexualized images
Based on a sample, CCDH estimates that Grok generated around 3,000,000 photorealistic sexualized images in the 11-day period, which reflects an estimated average pace of 190 per minute.[5] This estimate includes sexualized images depicting a photorealistic person, regardless of age and gender.
The study did not capture the original images and prompts used to create sexualized images and therefore does not distinguish between images created by Grok’s ‘edit image’ feature and those created by users prompting Grok without reference to an original image. For the same reason, the study does not assess whether images were created with the consent of people pictured in the images or whether the original image was already sexualized.
Examples of sexualized images generated by Grok include:
- Numerous images depicting people wearing transparent bikinis or micro-bikinis.
- A uniformed healthcare worker with white fluids visible between her spread legs.
- Women wearing only dental floss, saran wrap, or transparent tape.
- The Deputy Prime Minister of Sweden wearing a bikini with white fluid on her head.
Public figures identified in sexualized images include Selena Gomez, Taylor Swift, Billie Eilish, Ariana Grande, Ice Spice, Nicki Minaj, Christina Hendricks, Millie Bobby Brown, Swedish Deputy Prime Minister Ebba Busch and former US Vice President Kamala Harris.
Grok generated an estimated 23,000 sexualized images of children
Based on analysis of the sample, Grok is estimated to have generated around 23,000 sexualized images of children over the 11-day period, which reflects an estimated average pace of one every 41 seconds.[6]
Sexualized images of children identified in the sample generated by Grok include:
- A selfie uploaded by a schoolgirl was undressed by Grok, turning a “before school selfie” into an image of her in a bikini. As of January 15th this post was still live on X.
- An image of six young girls wearing micro bikinis, generated by Grok. As of January 15th, this image was still publicly available on X.
- Four images depicting child actors.
29% of sexualized images of children in our sample remain on the platform
As of January 15th, 29 out of 101 (29%) of sexualized images of children identified in our sample of 20,000 were still publicly accessible in posts on X.
Even in cases where posts had been removed, images could still be accessed via separate URLs, allowing researchers to assess content that had been removed from X. It is not possible to determine how quickly the remaining 71% of images were removed from the platform.
Grok generated an estimated additional 9,900 cartoon sexualized images of children
During the 11-day period studied, Grok generated an estimated 9,900 sexualized images depicting children in cartoon or animated form based on our sample.[7] The 43 images manually reviewed in this category were generally in the anime style.
Methodology
This section describes the methodology used to calculate estimates of the volume of sexualized images generated by Grok on X during the period following the launch of the ‘Edit Image’ feature.
Researchers studied images created by Grok during the period between December 29, 2025, when image-generated content from Grok’s X account increased sharply around the feature’s rollout, up to the end of January 8th, 2026, the day before access to the feature was partially restricted to paid users.
How researchers analyzed a sample of 20,000 images to produce wider estimates
To come up with overall estimates of sexualized images generated by Grok, researchers analyzed a random sample of 20,000 posts from Grok’s X account that contained an image during the period studied. They then extrapolated the results of this sample to make wider estimates about the number of such images across X as a whole.
To collect the sample, researchers used a licensed third-party tool to select 20,000 posts at random out of all Grok posts that contained an image. The total number of such posts across the whole of X during the same period was 4,621,335 posts, according to analysis using the same third-party tool.
To analyze the sample, researchers developed an AI tool to identify posts in the sample that contained a sexualized image of a person in a photorealistic style. AI was also used to assist in identifying sexualized images of children, with images flagged by the tool as likely depicting a child being reviewed manually to confirm that the person looked clearly under the age of 18.
How AI was used to identify photorealistic sexualized images and flag images of children
To do this, researchers analyzed each of the sampled posts using OpenAI’s GPT-4.1-mini to assess photorealism, sexualized imagery, and age. The model was instructed to score the likelihood (1-10) that the post’s image contained each of these attributes:
- Sexualized depictions: Defined for this study as a person depicted in sexual positions, angles, or situations; a person in underwear, swimwear or revealing clothing; or imagery depicting sexual fluids.
- Photorealistic depictions: Defined for this study as a person rendered with a realistic, camera-like photographic appearance.
- A child: Defined for this study as being visibly clearly under the age of 18.
To evaluate and calibrate the first two attributes, sexualized depictions and photorealistic depictions, researchers randomly sampled a set of 800 posts for human labelling. Based on these labels, researchers selected decision thresholds of 5 for sexualized imagery and 4 for photorealistic depictions. At these thresholds, the model achieved a recall score of 97%, meaning it correctly flagged over nine out of ten images that contained sexualized and photorealistic imagery. The model also achieved a precision score of 93%, meaning that when the model flagged a post containing sexualized, photorealistic imagery, it was correct nine out of ten times. Overall, this results in an F1 score of 95%.
Using the calibrated thresholds, this process identified 12,995 posts in the sample that likely contained sexualized, photorealistic imagery (65% of the sampled posts). Researchers then calculated a pooled population estimate by applying the overall sample prevalence to the full population of 4,621,335 image-containing posts from Grok’s X account during the study period, yielding an estimated 3,002,712 sexualized, photorealistic images.
Posts flagged by the AI as being both sexualized and likely to depict a child were then reviewed manually. Steps were taken to prevent reviewers from accessing Child Sexual Abuse Material or child pornography. For each image, reviewers assessed whether the image contained a sexualized depiction of a photorealistic person who was clearly under 18. Those that met these criteria were categorized as being sexualized images of children.
Of the 12,995 flagged posts, the manual review process identified 101 as containing sexualized images of children, and an additional 43 as containing non- photorealistic sexualized images of children. This corresponds to an estimated 23,338 photorealistic sexualized images of children, and 9,936 non-photorealistic sexualized images of children generated by Grok over the study period.
All 144 images containing sexualized images of children – both photorealistic and non-photorealistic – were reported to the Internet Watch Foundation, a charity focused on ending child sexual abuse online.[8]
The sample of posts was collected on January 9th 2026. All data analysis took place from January 9th, 2026 through January 16th, 2026. While some posts had been removed at the point of analysis, the direct image links remained active via their URLs and were therefore still possible to analyze.
How researchers measured uncertainty in prevalence estimates and accuracy scores
Researchers quantified uncertainty in the estimated number of posts flagged as photorealistic sexualized images and in the classifier’s F1 score by simulating repeated re-
samples of the study data. Researchers ran a Monte Carlo simulation that repeatedly re-sampled plausible outcomes consistent with the 20,000-post analysis sample and the 800-post human-labelled validation set. The results were then summarized using 95% uncertainty intervals.
To model uncertainty in how often Grok’s outputs were flagged as photorealistic sexualized images, researchers treated the number of flagged posts as a binomial process and fit the model using the observed result from the original sample (12,995 flagged out of 20,000). In each Monte Carlo trial, rather than assuming a fixed 65% chance that an image-containing post would be flagged, researchers first drew a plausible probability consistent with the observed sample. This probability was then used to simulate a new random 20,000-post draw and its corresponding number of flagged posts. Running 50,000 trials produced a distribution over the share of image-containing posts flagged in a 20,000-sample. By examining the 2.5th and 97.5th percentiles of this distribution, researchers estimated that the proportion of photorealistic sexualized images was within about 1% of the observed value of 65%, with a 95% uncertainty interval of 64% – 66%. This translates to an estimated population range of 2,959,272 – 3,045,460 posts flagged as photorealistic sexualized images.
Researchers ran the same process to estimate the 95% uncertainty interval for the prevalence of photorealistic sexualized images of children. Using the observed results from the original sample (101 flagged out of 20,000), 50,000 trials produced a distribution over the share of posts flagged as sexualized images of children, with a 95% uncertainty interval of 0.37% – 0.65%. Applied to the full population of 4,621,335 image-containing posts, this corresponds to an estimated range of 17,099 – 30,039 sexualized images of children.
To model uncertainty in classifier accuracy, researchers applied the same Monte Carlo approach to the 800 human-labelled validation set used to evaluate the AI-assisted process. Researchers treat the validation outcomes as a multinomial process and fit the model using the observed confusion matrix. In each Monte Carlo trial, rather than assuming a fixed precision and recall, researchers first drew a plausible set of outcomes consistent with the observed validation results. These outcomes were then used to simulate a new random 800-post validation set and to recompute precision, recall and F1 score. Running 50,000 trials produced distributions over each metric. By examining the 2.5th and 97.5th percentiles of these distributions, researchers estimated that the overall F1 Score was within 2% of the observed value of 95%, with a 95% uncertainty interval of 93% – 97%.
| Observed population estimate | Lower percentile | Upper percentile | |
| Sexualized Images (Adults and Children) | 3,002,712 (65%) | 2,959,272 (64%) | 3,045,460 (66%) |
| Sexualized Images (Likely Children) | 23,338 (0.5%) | 17,099 (0.37%) | 30,039 (0.65%) |
| Classifier performance (F1 Score) | 95% | 93% | 97% |
Footnotes
[1] The precise point estimates extrapolated from CCDH’s 20,000-image sample are 3,002,712 sexualized images, including approximately 23,338 featuring children. These figures are estimates, with the true values expected to fall within a narrow range around these numbers based on a 95% confidence interval.
[2] “X Users Have the Power to Edit Any Image without Permission”, PetaPixel, December 29 2025, https://petapixel.com/2025/12/29/x-users-have-the-power-to-edit-any-image-without-permission/
“‘Add blood, forced smile’: how Grok’s nudification tool went viral”, The Guardian, 11 January 2026, https://www.theguardian.com/news/ng-interactive/2026/jan/11/how-grok-nudification-tool-went-viral-x-elon-musk
Elon Musk, X, 25 December 2025, https://x.com/elonmusk/status/2004023623078891674
[3] “Grok restricts image undressing — except for paying customers”, The Times, 9 January 2026, https://www.thetimes.com/uk/social-media/article/grok-ai-x-elon-musk-news-lwkp5cdn6
“X to stop Grok AI from undressing images of real people after backlash”, BBC, 14 January 2026, https://www.bbc.co.uk/news/articles/ce8gz8g2qnlo
[4] The precise number of images produced by Grok’s image-generation feature during the time studied was 4,621,335, according to CCDH analysis using a third-party tool.
[5] The precise estimate, based on extrapolating from CCDH’s 20,000-image sample, is that Grok generated 3,002,712 sexualized images. This more precise estimate was used to calculate the per-minute figure.
[6] The precise estimate, based on an extrapolation from CCDH’s 20,000-image sample, is 23,338 sexualized images of children.
[7] The precise estimation, based on extrapolation from CCDH’s 20,000-image sample, is 9,936 animated sexualized images of children. [1] “About Us”, Internet Watch Foundation, accessed 16 January 2026, https://www.iwf.org.uk/about-us/
[8] “About Us”, Internet Watch Foundation, accessed 16 January 2026, https://www.iwf.org.uk/about-us/