Author: admin

The UK technology secretary has called a wave of images of women and children with their clothes digitally removed generated by Elon Musk’s Grok AI “appalling and unacceptable in decent society”.

After thousands of intimate deepfakes circulated online, Liz Kendall said X, Musk’s social media platform, needed to “deal with this urgently” and she backed the UK regulator Ofcom to “take any enforcement action it deems necessary”.

“We cannot and will not allow the proliferation of these demeaning and degrading images, which are disproportionately aimed at women and girls,” she said. “Make no mistake, the UK will not tolerate the endless proliferation of disgusting and abusive material online. We must all come together to stamp it out.”

Her comments came amid warnings that the Online Safety Act, which aims to tackle online harms and protect children, needs to be urgently toughened up despite pressure from the Trump administration to water it down.

One expert criticised the “tennis game” between platforms such as X and UK regulators when problems arose and called the government response “worryingly slow”.

Jessaline Caine, a survivor of child sexual abuse, called the government’s response “spineless” and told the Guardian that on Tuesday morning the chatbot was still obeying requests to manipulate an image of her as a three-year-old to dress her in a string bikini. Her identical requests made to ChatGPT and Gemini were rejected.

“Other platforms have these safeguards so why does Grok allow the creation of these images?” she said. “The images I’ve seen are so vile and degrading. The government has been very reactive. These AI tools need better regulation.”

On Monday, Ofcom said it was aware of serious concerns raised about Grok creating undressed images of people and sexualised images of children. It said it had contacted X and xAI “to understand what steps they have taken to comply with their legal duties to protect users in the UK” and would assess the need for an investigation based on the company’s response.

The pressure is growing on ministers to take a tougher line. The crossbench peer and online child safety campaigner Beeban Kidron has urged the government to “show some backbone” and called for the Online Safety Act regime to be “reassessed so it is swifter and has more teeth”.

Speaking about X, she said: “If any other consumer product caused this level of harm, it would already have been recalled.”

She said Ofcom needed to act “in days not years” and called for users to walk “away from products that show no serious intent to prevent harm to children, women and democracy”.

Ofcom has powers to fine tech platforms up to £18m or 10% of their qualifying global revenues, whichever is higher. The biggest penalty to date came last month when a porn provider that failed to carry out mandatory age checks was fined £1m.

Last month, ministers promised new laws to ban “nudification” tools, which use generative AI to turn images of real people into fake nude pictures and videos without their permission. It remains unclear when that ban will be enforced.

Sarah Smith, the innovation lead at the Lucy Faithfull Foundation, a charity that works to prevent child abuse, called for X to immediately disable Grok’s image-editing features “until robust safeguards are in place to stop this from happening again”.

X did not respond to a request for comment on Kendall’s remarks. It said on Monday: “We take action against illegal content on X, including child sexual abuse material, by removing it, permanently suspending accounts and working with local governments and law enforcement as necessary.”

Jake Moore, a global cybersecurity adviser at at the security software firm ESET, criticised the “tennis game” between platforms such as X and UK regulators and called the government response “worryingly slow”.

He said that as AI increasingly allowed faked images to be rendered as longer videos, the consequences for people’s lives would only become worse.

“It is unbelievable that this is able to occur in 2026,” he said. “We have to move forward with extreme regulation. Any grey area we offer will be abused. The government is not understanding the bigger picture here.”

It is already illegal to create or share non-consensual intimate images or child sexual abuse material, including sexual deepfakes created with AI. Fake images of people in bikinis may qualify as intimate images, as the definition in law includes the person having naked breasts, buttocks or genitals or having those parts only covered by underwear. Indecent images include those depicting children in erotic poses without sexual activity.

Lady Kidron said AI-generated pictures of children in bikinis may not be child sexual abuse material but they were contemptuous of children’s privacy and agency.

“We cannot live in a world in which a kid can’t post a picture of winning a race unless they are willing to be sexualised and humiliated,” she said.

We aimed to quantify the relationship between the response elicited by the bi-speed stimuli and the corresponding component responses. We first assumed that the response R of a neuron elicited by two component speeds can be described as a weighted sum of the component responses R_s and R_f elicited by the slower ($v_s$) and faster ($v_f$) component speed, respectively Equation 1.

(1)

${\displaystyle {\displaystyle R\left(v_s,v_f\right)=w_s\left(v_s,v_f\right)\bullet R_s+w_f\left(v_s,v_f\right)\bullet R_f,}}$

in which, w_s and w_f are the response weights for the slower and faster speed component $v_{s}and{v}_f$, respectively.

Our goal was to estimate the weights for each speed pair and determine whether the weights change with the stimulus speeds. In our main data set, the two speed components moved in the same direction. To determine the weights of $w_s$ and w_f for each neuron at each speed pair, we have three data points R, R_s, and R_f, which are trial-averaged responses. Since it is not possible to solve for both variables, $w_s$ and w_f, from a single equation Equation 1 with three data points, we introduced an additional constraint: $w_s$ + w_f = 1. With this constraint, the weighted sum becomes a weighted average. While this constraint may not yield the exact weights that would be obtained with a fully determined system, it nevertheless allows us to characterize how the relative weights vary with stimulus speed. As long as R_f ≠ R_s, R can be expressed as:

(2)

${\displaystyle {\displaystyle R=\frac{R_f-R}{R_f-R_s}{R}_s+\frac{R-R_s}{R_f-R_s}{R}_f,}}$

The response weights are $w_s=\frac{R_f-R}{R_f-R_s}$ , $w_f=\frac{R-R_s}{R_f-R_s}$. Intuitively, if R were closer to one component response, that stimulus component would have a higher weight. Note that Equation 2 is not intended for fitting the response R using $R_s$ and $R_f$, but rather to use the relationship among R, $R_s$, and $R_f$ to determine the weights for the faster and slower components.

Using this approach to estimate response weights for individual neurons can be unreliable, particularly when R_f and Rs are similar. This situation often arises when the two speeds fall on opposite sides of the neuron’s preferred speed, resulting in a small denominator (R_f – Rs) and consequently an artificially inflated weight estimate. We, therefore, used the neuronal responses across the population to determine the response weights (Figure 5). For each pair of stimulus speeds, we plotted (R−R_s) in the ordinate versus (R_f − R_s) in the abscissa. Figure 5A1–E1 shows the results obtained at 4x speed separation. Across the neuronal population, the relationship between (R – R_s) and (R_f − R_s) can be described by a linear equation (Equation 3) (see R² in Table 1). This linearity suggests that the response weights for each speed pair are roughly consistent across the neuronal population.

(3)

${\displaystyle {\displaystyle R-R_s=k\left(R_f-R_s\right)+b}}$

Because all the regression lines in Figure 5 nearly go through the origin (i.e. intercept b ≈ 0, Table 1), the slope k obtained from the linear regression approximates $\frac{R-R_s}{R_f-R_s}$, which is the response weight $w_f$ for the faster component (Equation 2). Hence, for each pair of stimulus speeds, we can estimate the response weight for the faster component using the slope of the linear regression of the responses from the neuronal population.

Our results showed that the bi-speed response showed a strong bias toward the faster component when the speeds were slow and changed progressively from a scheme of ‘faster-component-take-all’ to ‘response-averaging’ as the speeds of the two stimulus components increased (Figure 5F1). We found similar results when the speed separation between the stimulus components was small (2x), although the bias toward the faster component at low stimulus speeds was not as strong as 4x speed separation (Figure 5A2–F2 and Table 1).

In the regression between $\left(R{-R}_s\right)$ and $\left(R_f-R_s\right)$, $R_s$ (i.e. the firing rate to the slow component averaged across all trials for each neuron) was a common term and, therefore, could artificially introduce correlations. We wanted to determine whether our estimates of the regression slope ($w_f$) were confounded by this factor. We performed two additional analyses.

First, at each speed pair and for each of the 100 neurons in the data sample shown in Figure 5, we simulated the response to the bi-speed stimuli ($R_e$) as a randomly weighted average of $R_f$ and $R_s$ of the same neuron.

(4)

${\displaystyle {\displaystyle R_e=a{R}_f+\left(1-a\right)R_s,}}$

in which $a$ was a randomly generated weight (between 0 and 1) for $R_f$, and the weights for $R_f$ and $R_s$ summed to one. We then calculated the regression slope and the correlation coefficient between the simulated $R_e-R_s$ and $R_f-R_s$ across the 100 neurons. We repeated the process 1000 times and obtained the mean and 95% confidence interval (CI) of the regression slope and the R². The mean slope based on the simulated responses was 0.5 across all speed pairs. The estimated slope ($w_f$) from the data was significantly greater than the simulated slope at slow speeds of 1.25/5, 2.5/10 (Figure 5F1), and 1.25/2.5, 2.5/5, and 5/10°/s (Figure 5F2) (bootstrap test, see p-values in Table 1). The estimated R² based on the data was also significantly higher than the simulated R² for most of the speed pairs (Table 1).

Second, we calculated $R_s$ in the ordinate and abscissa of Figure 5A–E using responses averaged across different subsets of trials, such that $R_s$ was no longer a common term in the ordinate and abscissa. For each neuron, we determined $R_{s1}$ by averaging the firing rates of $R_s$ across half of the recorded trials, selected randomly. We also determined $R_{s2}$ by averaging the firing rates of $R_s$ across the rest of the trials. We regressed $\left(R{-R}_{s1}\right)$ on $\left(R_f{-R}_{s2}\right)$, as well as $\left(R{-R}_{s2}\right)$ on $\left(R_f{-R}_{s1}\right)$, and repeated the procedure 50 times. The averaged slopes obtained with $R_s$ from the split trials showed the same pattern as those using $R_s$ from all trials (Table 1 and Appendix 1—figure 1), although the coefficient of determination was slightly reduced (Table 1). For 4x speed separation, the slopes were nearly identical to those shown in Figure 5F1. For 2x speed separation, the slopes were slightly smaller than those in Figure 5F2, but followed the same pattern (Appendix 1—figure 1). Together, these analysis results confirmed the faster-speed bias at the slow stimulus speeds and the change of the response weights as stimulus speeds increased.