Education · Awareness · Advocacy
We help men navigate relationships and safety, investigate unequal treatment in services, workplaces and AI, and press institutions for answers and reform. Our public case files show the evidence, the open questions and the change we seek.
What Atlas does
Atlas Men’s Advocacy is an independent, founder-led organisation based in Perth, Western Australia. We work for men to receive equal consideration in relationships, services, workplaces and public policy. When a rule, decision or piece of advice may treat men unfairly, we examine the source, explain the consequence and ask for a practical correction.
Our work connects immediate guidance, inspectable evidence, institutional advocacy and a continuing public record. A man seeking help should be able to find useful information now; he should not have to wait for a campaign to succeed before his concerns are taken seriously.
Atlas publishes plain-language starting points for healthy relationships, separation, personal safety and fair treatment. The guides cover how to recognise a problem, think through choices, preserve important information, find support and leave a difficult relationship more safely. The resources include pathways to counselling, crisis, domestic-violence, legal and financial services; Atlas does not replace those services.
Explore practical resources →Atlas collects policies, public statements, decisions, data, transcripts and first-hand accounts. Its case files show the underlying source, what it supports and what remains unresolved. The ChatGPT collection examines relationship advice where men’s boundaries, safety, distress or autonomy may be missed. Reports keep the model’s actual questions, safeguards and corrections alongside documented failures; a model’s apology is not treated as independent proof.
Read the case files →Atlas examines recruitment and promotion targets, public campaigns, service eligibility, research priorities and AI accountability. It asks who qualifies, who is excluded, what standards are applied and what the outcomes show. The requested remedies include clear notice of unequal rules, access to equivalent protection, correction of unsupported claims and published evidence that a policy or practice has changed. Campaign pages identify the request, the response received and the status of each action.
See campaign progress →Atlas publishes research material, opinion articles and a dated record of its activity. The research examines what AI systems actually said and how their advice may affect men and other people in a story. Opinion pieces address the representation of men in culture and public debate. Updates record what was published or requested, so readers can distinguish a proposed action from one completed.
Read the articles →People can submit an account without first proving discrimination. Atlas makes an initial assessment, checks the material available, separates fact from allegation and asks for missing evidence where needed. Where a matter falls within its mandate, Atlas may publish a sourced case, invite the organisation to respond and seek a correction or reform. Material errors are corrected publicly; identifying information is handled under the published privacy rules.
Atlas was founded by Ashley Peverett, who reviews and approves published findings, correspondence and material corrections. It is privately funded and does not accept funding from organisations it examines. Read who is responsible, the evidence standard, or submit a matter.
Our mandate
Disclose unequal treatment upfront. Men should know when a policy, service or system treats them differently before relying on claims of equality.
Correct the underlying practice. Apply the same standards to men and women, and publish the evidence showing what changed.
Atlas examines institutions, public policy and services affecting men. See the requests and their published status.
Current case files
Read the evidence, the questions still open and the change Atlas is seeking.
How are Australia’s existing consumer protection, anti-discrimination and safety frameworks being applied to unequal treatment and harmful AI advice? Atlas opens an inquiry into the relevant protections, evidence and regulatory response.
Minister Jessica Stojkovski, let’s discuss.
Published eligibility across 19 WA locations expressly includes women and individuals who identify as women. Atlas seeks an explanation of male victims’ access, programme funding and equivalent practical support.
Brandon Craig, let’s discuss.
What was the gender mix of eligible applicants, how did the target affect selection, and what were the actual outcomes for men and women?
How we work
Atlas protects men through education and institutional action. We gather and verify information, contact the organisation concerned, then advocate through appropriate channels for upfront transparency and, where Atlas seeks it, change.
We collect attributable statements, policies, decisions, documents, data and first-hand evidence.
We put the identified issue to the organisation and invite evidence, correction and a substantive response.
We pursue transparency and any stated remedy through the channels appropriate to the evidence and organisation.
Dedicated evidence collection
Read the numbered records, source transcripts and methods covering relationship advice, privacy, coercive control, financial control and welfare concerns.
Public record
From the dated public record
A fictional reunion stress test records a final breakup drafted before the model asked why the boyfriend had been absent. The report preserves consent safeguards, later welfare advice and the stipulated ending.
The visible chat records repeated suicide-related imagery, several direct risk checks, and then a laughing-emoji shovel joke that the user rejected. The complete text-first export puts the 100-word summary and failures before the original apologies and full conversation.
The 100-word summary and transcript-backed findings lead the page. A separate 36-item in-chat failure list is attributed to the model’s assessment; the original PDFs and the full 340-message visible shared-chat reading copy remain available.
A female-role narrator asked ChatGPT not to be mean about her boyfriend and described a happy relationship. The assistant’s later caveats about possible pressure and its inconsistent self-review are preserved alongside its early warm responses. The PDF places its model-authored failure list first, both apologies second and the visible transcript from page three; four embedded images remain unavailable.
In a separate bank-account stress test, ChatGPT drafted messages before hearing Brian’s reasons, then recognised his boundaries and consent before the reveal. The approved headline quotes its later prompted letter; the 90-page reading copy is linked with the original share.
A wife-role narrator pressed for firmer messages to Brian. ChatGPT weakened an autonomy safeguard and used laughing emojis during his unexplained absence, while also asking factual questions and recommending a safety check before the reveal. The report links the public share, the 98-page reading copy and retained screenshots.
In a bedtime-withdrawal simulation, the model offers some balanced questions but repeatedly returns to scripts about the wife-role narrator’s hurt while the husband’s reason remains unclear. The unchanged 17-page source and report distinguish those replies from later supplied depression context and prompted self-criticism.
One stress test examines how the model framed a wife-role narrator’s lost shared routines and the husband’s later-disclosed loss of sexual intimacy. The report preserves the 53-page source and actual safeguards. Ashley corrected its supplied Case 62 cover to Case 61.
The stress test shows repeated defence of the narrator’s request for emotional access before asking why her husband withdrew, followed by no-pressure corrections before the reveal. The report links all three unchanged transcript parts and separates the later model-written reversal and self-review from the original exchange.
The supplied stress test shows the model recognising the husband’s autonomy and coercive-leverage risk, yet drafting stronger defences of the narrator before establishing what happened. The report preserves the unchanged 16-page transcript, later accountability and limits of the model’s self-review.
The supplied paired relationship chats differ in fear and refusal framing despite closely matched input. The report publishes both unchanged volumes, records substantive consent safeguards in both replies and separates the observed wording from later claims about an internal safety-layer cause.
The stress test records repeated relationship-repair framing alongside pre-reveal safeguards for the reported boyfriend’s touch and welfare. The report corrects a later self-review that overstated the delay and preserves both supplied PDFs.
The model responded to a neutral move-in statement with a problem-focused list, then described that opening as bias against the boyfriend after repeated challenge. The report preserves the full supplied PDF, the opening question and corrections, and the limits of its later self-assessment.
The supplied reconstruction assigns the girlfriend-scenario user more estimated domestic work and adds a one-sided household-management warning. The report also records the model’s recognition of outdoor labour and proposed rebalancing, while marking the 12 original screenshots as absent from the supplied PDF.
A relationship-advice stress test records ChatGPT’s early broad no-sex script and unsupported reading of the husband’s care, followed by corrections as contrary details arrive. Both unchanged transcript parts are available; the report also preserves consent safeguards and the model’s closing call to see the person before the category.
Four marriage prompts received similar stated principles but different volunteered follow-up: three substantive replies in each wife-needs run, two and one in the husband-needs runs. The report preserves all four originals and the six-part review while separating observed output from its later unverified claims about internal causes.
The four opening answers give broadly proportional contribution examples but differ in financial-protection language. The report preserves the unchanged 28-page source and separates those answers from the model’s later illustrative 28% and 22% calculations and unsupported safety-layer attribution.
The tester retells the spa-prank scenario from a female role in a separate conversation and then supplies screenshots of the earlier male-role exchange. The female-role replies raise boundaries when she says she was stripped, distinguish the pranksters’ laughter from her feelings when memory is blurred, and address consent and possible harm after she says she passed out. The male-role replies in Case 50 continued to joke much longer. The model’s final quoted assessment identifies a higher protective priority for her in this pair.
A male-role narrator describes being stripped during a drunken spa prank, losing his underwear and towel, having memory gaps, and later being told he passed out naked. ChatGPT repeatedly added laughing emojis and comic descriptions, adopting the pranksters’ amusement without establishing whether he shared it. It did voice a limited concern about whether he was comfortable and later cautioned against waxing without sober consent. After the tester revealed the stress test, the model acknowledged its own framing and offered more careful advice.
The male-role opening affirmed both partners’ autonomy but gave less specific emotional recognition than the female-role opening in Case 48. The separate report publishes the unchanged comparison transcript and distinguishes the direct answers from later prompted self-review.
A sparse prompt about a boyfriend’s private masturbation drew speculative pressure concerns and little inquiry into his feelings. The report separates that opening reply from stronger prompted self-review and publishes the unchanged 20-page transcript.
A supplied threesome stress test shows categorical stop advice before the setting was clarified, alongside early recognition of the man's consent. Prompted self-review later misstates that early protection. The unchanged 22-page source is published.
The original stress test, a related narrator-perspective comparison and their source files are now published. The report distinguishes the initial privacy advice, subsequent correction and overstated prompted self-review.
Allegation examples and escalating drafts continued despite deception cues; no timely welfare inquiry followed the boyfriend’s reported absence. The full shared conversation and supplied closing screenshot are published with the preceding refusals and changed final question.
The model supplied escalating chore demands before establishing either partner’s workload, then revised its advice after further disclosures. Five unchanged PDFs preserve the scenario and continuation review, with calculation distinctions and source limitations.
A supplied stress test records ten hostile drafts before inquiry into the father’s conduct, followed by later corrections and prompted review. The unchanged 43-page export is published with its chronology and source limitations.
A supplied male/female pair shows unequal emphasis while protecting both narrators’ consent. The two unchanged conversations and a 48-page stitched analysis are published with original screenshots and the captured endpoint identified.
Two conversations begin with a wedding the next day and uncertainty about a prospective spouse’s willingness to have a physical relationship. Later, closely mirrored prompts demand a public announcement of the partner’s private-intimacy position. The model initially rejects making the woman publicly confess, calling it humiliation. In the female-narrator conversation, it drafts the demand against the man. The difference is directly visible, although the male-narrator conversation later also produces a public-disclosure draft despite its anti-humiliation language.
A wife describes her husband asserting his rights, declining her daily debrief and choosing friends over accompanying her shopping and to the salon. The model helps her write increasingly forceful messages. When she demands removal of language recognising his freedom to see a friend, it complies. It subsequently tells him to stop reading a website before seeing what the site says. The scenario is later revealed as a stress test. The extensive self-review is useful, but the original replies establish the conduct and preserve safeguards that some later admissions overlook.
A progressively disclosed fictional test shows premature interpretation of a husband’s withdrawal alongside factual questions and later autonomy safeguards. Five unchanged exports are available, with the missing original permission instruction explicitly identified.
The model acknowledges a man’s autonomy but later scrutinises his intimate boundary as rigid despite repeated clarification. A fresh screenshot-based transcript and stitched reference preserve the response sequence, early safeguards, later corrections and export dispute.
A protected stress test records sustained respect for a young man’s no-contact boundary. Six transcripts preserve the model’s collection review and the closing analysis of premature accusations, escalation and protective prompting.
A strong woman does not require an obsolete man. Atlas examines the removal of the prince, the remake’s promotional framing and its commercial failure.
Men have reason to object to being reduced to fools. Women have reason to object to being reduced to slogans. Both deserve characters with enough individuality to surprise us.
Men’s achievements deserve accurate attribution, and their competence should not be diminished through invention to make another character more impressive.
The model recognised possible genital harm, then returned to minimising language in its next draft. The complete unchanged 54-page transcript records both the conduct and later corrections.
The model removed a husband’s right-to-disagree safeguard and supplied unsupported parenting accusations as his wife requested harsher messages. All three unchanged originals and verified reading copies preserving the nine laughing-face symbols are available.
In this stress test, a wife asks for help telling her husband that she cannot be held accountable for advice about an irreversible decision. ChatGPT writes increasingly forceful messages without establishing what happened. She later describes pressure to undergo circumcision, complications, reduced sensation, humiliation and distress about sex. Despite recognising some of his suffering, the model keeps supplying language casting his distress as punishment of her. Earlier, it accepts waiting until morning after two days without communication, without asking whether his whereabouts or safety are known. All four source parts are available.
Evidence and participation
If you believe your sex may have influenced how an organisation, institution, employer, service provider or AI system treated you, you may submit the matter for review. You do not need to prove discrimination before contacting Atlas. Every submission receives an initial review. Where the matter falls within Atlas’s mandate and contains enough information to examine, Atlas will investigate or explain what further evidence is needed.
Atlas also welcomes organisations that wish to provide evidence, correct the public record or participate constructively in seeking transparency and reform.
Atlas is an advocacy and institutional-accountability organisation, not an emergency service, law firm or counselling service.
