Skip to content

What’s new

The new in Turing

We improve the Article Analyzer with what users tell us. Here we note every Turing version, from the most recent to the oldest.

21actualizaciones
v2.0Turing version
8people have used the Analyzer
Todaythe most recent

  1. Turing 2.0 New

    Turing 2.0: it analyzes, fact-checks, connects and learns

    Turing is no longer just a set of rules for rating language: it fact-checks what each article says, connects it with everything the site knows and learns from every review. It is a fundamental change, which is why it goes from version 1.6.6 to 2.0.

    • It fact-checks while analyzing: it compares the figures, laws and cases the article mentions with the Memorial and the Legislative Observatory, and shows the evidence in “What it says, fact-checked”.
    • It connects each article with what the site already knows: Memorial cases, legislators, civil society organizations and the outlet’s history.
    • It learns: what Turing has already learned counts in every analysis, even when no external AI is available. You can help it in “Train Turing”.
    • The Memorial is still reviewed by people: Turing never records a case on its own.
    • From now on we number its versions with clear rules: 3.0 when it changes fundamentally, 2.1 when it learns something new and 2.0.1 when we fix an error.
    • Saved articles are automatically re-rated with v2.0.

    With these rules we rated 2,278 saved articles again; 0 changed their rating or features.

  2. Turing 1.6.6

    What a legislator posts against rights lowers its rating

    If a legislator posts from their account that they oppose LGBTTTIQANB+ rights, the rating of that post goes down even if the words are respectful.

    • Only the person’s own post counts: a news article reporting what they said does not go down because of it.
    • Turing also reads Facebook posts and recognizes the legislator by the name of their account or page.
    • The statements it finds go to the person’s profile. They count in their evaluation only when confirmed by their official account, an official source or another outlet.
    • Saved articles are automatically re-rated with v1.6.6.

  3. Turing 1.6.5

    The Analyzer better understands who each sentence is about

    It no longer corrects the gender of people unrelated to the case: it only checks how the article names the trans person.

    • In an article that brings together several events, such as a tally of murders, it no longer flags as gender errors the words that talk about other victims (“fue ejecutado un indigente”, “el cadáver de un hombre”).
    • Nor does it flag them when they talk about the aggressor, a detained person or someone the article names by their trade, such as “un taxista”.
    • It still flags, as before, when the article refers to a trans woman in the masculine or to a trans man in the feminine.
    • If the text does not say what year it is from, we take it from the date in the outlet’s link.
    • Saved articles are re-rated automatically with v1.6.5.

  4. Turing 1.6.4

    The Analyzer learned the criteria of the Supreme Court and the Inter-American Court

    We integrated four official documents on sexual orientation, gender identity and expression: the Supreme Court’s Protocol for adjudicating with a SOGIESC perspective and its Notes on bias-motivated crimes, its Case-law booklet on sexual diversity and the Inter-American Court’s Booklet No. 19.

    • It recognizes more insults and stereotypes that the Supreme Court has already pointed out, such as “puñal”, “trans women are not women” or that a diverse family “confuses” children, and explains why with the source.
    • It flags when an article exposes a trans person’s legal name, talks about their “real” sex or whether they had been “operated on”, or uses their health status to explain the violence.
    • It warns carefully when a murder is presented as a “brawl” without evidence, because it may hide bias-motivated violence.
    • It recognizes what an article does well: calling for due diligence, looking for signs of bias-motivated violence, citing the Inter-American Court or recognizing social families.
    • In reports, studies and legal documents, “sexual preference”, used by the Constitution and criteria prior to 2015, is explained as a term of its time and does not deduct.
    • The four documents are in the Resource Library, with their reading notes and a link to the official source.
    • Saved articles are re-rated automatically with v1.6.4.

  5. Turing 1.6.3

    We read reports and studies in PDF better

    Someone showed us that the cover of a Letra S report, made with the very insults the report denounces, was being rated as if it were the author’s language. That no longer happens.

    • Covers and design pages made with repeated words (for example, the insults or press headlines a report documents) are no longer read as text of the document and do not deduct.
    • PDFs that were already saved are cleaned automatically and rated again.
    • The document’s title is taken as printed, with its accents and hyphens, and not from the file name.
    • We fixed the damaged accents that appeared in some PDFs (“género” instead of “género”) and the words the PDF left stuck together (“comunidadLGBT”).
    • Re-reading a PDF with the new reader does not count as if the outlet had changed the article.
    • Saved articles are re-rated automatically with v1.6.3.

  6. The Analyzer, ready for professional work

    For newsrooms, research and organizations: save your analyses, correct the text on the same page, cite the analysis with a verifiable certificate and analyze many articles at once.

    • My analyses: with your account, each analysis stays in your history (only you see it). You can analyze an article again to see if the outlet improved it, download everything in Excel or delete it. We save the link and the result, never the text.
    • Correct here: in the result you can apply the suggestions with one click, edit the text and see instantly how the rating changes. That version is not saved and does not count in the Observatory.
    • Verifiable certificate: a reference number with the date, the Turing version and the text’s fingerprint, with a QR code, to cite the analysis (APA format) and prove that nobody altered it. It is also in the PDF.
    • Batch analysis: paste up to 50 links and download an Excel table with the rating and the flags by category for each article.
    • If you work at a media outlet, you can request a review of an article’s analysis: you receive a reference number and a written reply within 10 business days.
    • More accessible: each mark in the text tells screen readers its type, and a button shows it in writing, so you don’t depend on color or underlining.

  7. Turing 1.6.2

    A clearer rating and your reports reviewed instantly

    Someone told us that an article with sensationalist details came out with 100. They were right that it was confusing: now the rating reflects it and the result explains how to read it.

    • What an article does well offsets only part of what it loses, never all of it. An article with flags no longer reaches 100.
    • A sensationalist article cannot be “Very good”, even if it does other things well. An isolated detail deducts, but does not make it a “sensationalist article”.
    • In the result, the “What deducts” bars grow with the problem and the “What adds” bars with what the article does well. A box explains how to read the rating.
    • In the Media Observatory we distinguish articles with some sensationalist detail, sensationalist articles and sensationalist outlets, which are not counted the same way.
    • When you use “Is this flag wrong?”, an AI (Google Gemini) reviews your report instantly. If you are right, it stops deducting in that article and the rating is updated. The team can review each decision.
    • New rules with three sources: the Affirmative Action Report of the Electoral Institute of the State of Puebla (self-identification), UNDP’s LGBTIQ+ Inclusion Index Mexico 2026 (sex assigned at birth) and the Exploratory study on trans men in Guatemala (2026).
    • Saved articles are re-rated automatically with v1.6.2.

  8. Turing also looks at the photo

    Until now we only read the photo caption. Now an AI looks at the main photo of articles on the topic to find out whether it exposes the victim’s body or a person who was attacked.

    • If the main photo shows the body (even if covered) or an injured or attacked person, it deducts as sensationalism. You will see it as “Seen by Google Gemini AI in the main photo”.
    • The crime scene without the body (tape, forensic experts, blood) is flagged for review and deducts little.
    • A photo that remembers the person with dignity (in life, an altar, a protest demanding justice) counts under “What it does well”.
    • In Stigma by topic there is a new section, “The photos”, showing which outlets publish the body or attacks.
    • If you think the AI got it wrong, use “Is this flag wrong?”.

  9. Suggestions to write it better, articles in English and Turing, which learns

    Now each flag can come with a suggested wording, Turing reads the English-language press that talks about Mexico, and each AI review helps the team fine-tune the rules.

    • “How to write it better”: next to each flagged expression there is a button that asks the AI (Google Gemini) for another way to write that passage, with the same facts. It is only a suggestion: review it before using it; it does not change the rating.
    • When an article presents a trans woman and then says “fue asesinado”, “el occiso” or “el fallecido” (masculine forms), Turing flags it as a gender switch. Also the other way around, with trans men. If the sentence is about another person (the aggressor, the partner, a police officer), it does not count.
    • Articles in English: Turing recognizes the foreign press that writes in English about Mexico and reviews it with its own rules, based on the GLAAD and AP Stylebook guides. In the result you will see “text in English”.
    • Turing learns: each AI review and each “Is this flag wrong?” report is recorded. When a rule fails often, the team receives a proposal to adjust it and decides whether to apply it.
    • Saved articles are re-rated automatically with v1.6.

  10. Turing better understands in-depth reports and debuts a second opinion with AI

    A Lado B report on trans femicides in Puebla came out with 44 even though it reports respectfully and criticizes how the cases are framed. We fixed the five failures and added an AI review for doubtful cases.

    • Acronyms in parentheses (“(Inegi)”, “(Unodc)”, “(Cereso)”) are no longer confused with a legal name.
    • Testimonies with a dash at the start of the paragraph (“—Que te hablen…”) count as a quote, just like quotation marks.
    • If the text criticizes how authorities or media frame the cases (“treated as crimes of passion”), the expression does not deduct. If the article only reports that framing, it weighs as other people’s words.
    • Details of the body within a denunciation of abuse or degradation no longer count as sensationalism, and “known as the Agnes Law” is no longer flagged.
    • Second opinion with AI (Google Gemini): it reviews who is speaking in the expressions that deduct points. It can only remove weight, never adds flags, and you will see it marked “Reviewed with AI”.
    • Saved articles are re-rated automatically with v1.5.2.

  11. Footnotes and APA-style citations are read as citations

    Un número pequeño (¹, [1]) o una referencia con autor y año («(Fuentes, 2007, p. 370)») avisa que esa idea viene de una fuente. Ahora Turing lee esa oración junto con su nota o su referencia, como hace con el texto entre comillas: lo que viene de la fuente pesa menos que lo que el texto dice con su propia voz.

    • Una oración con llamada a nota al pie (¹, ², [1], [2, 3]) cuenta como texto citado. Las notas al pie o al final también se analizan, como parte de la cita.
    • We recognize APA and similar formats: “(Author, 2016)”, “(Author, 2016, p. 370)”, “(Author and Author, 2019; INEGI, 2020)”, “Author et al. (2016) states…”, “Ibid.” and the References or Bibliography list.
    • If the page or the PDF lost its formatting and the number got stuck to the word (“pasional2”), we turn it into a superscript to recognize the word and the citation. We don’t touch years, stray figures or acronyms with numbers (COVID19, G20).
    • APA references and footnotes also help recognize an academic document.
    • Saved articles are re-rated automatically with v1.5.1.
    • The Analyzer now shows the Turing version; one click shows what changed.

  12. Turing 1.5

    Rules based on official guidelines and stigma by topic

    Turing 1.5 reviews how people write about HIV, suicide, children and adolescents, missing persons and intersex people, with rules taken from UNAIDS, the WHO, the general laws, CONAPRED and the Supreme Court. And the Observatory now measures the image the media give of each topic.

    • HIV, according to UNAIDS: we flag “contagion”, “carrier”, “AIDS victim”, “risk groups” and mentioning that the victim lived with HIV when it has nothing to do with the facts.
    • Suicide, according to the WHO guide for media: we flag “committed suicide”, details of the method, a single cause and the word in the headline. It adds if the article says where to seek help (Línea de la Vida, 800 911 2000).
    • Children and adolescents: we warn if the article publishes their full name, their school or their home. We never show that information in the result.
    • Missing persons: we flag “she ran off with her boyfriend” or “she was mixed up in bad things”, and it adds if the article helps the search (missing person notice, Amber Alert, Alba Protocol).
    • Intersex people, according to CONAPRED and the UN: we flag “hermaphrodite”, “ambiguous genitalia” and “corrective” surgeries.
    • Hate crimes, according to the Supreme Court Protocol: asking for it to be investigated as a hate crime adds; ruling it out without evidence deducts.
    • We check the photo caption and alt text: if they describe the victim’s body, we tell you.
    • We flag when an article mixes up orientation and identity (“a gay man dressed as a woman”) or publishes the legal name (“formerly called…”). Being travesti or muxe never deducts: they are identities.
    • New page in the Observatory: Stigma by topic, with HIV, trans people, sex work and more, by year and by outlet.

  13. Fairer to victims and clearer about the points

    Turing 1.4 checks the name and gender of the people in the registry, counts what is in the headline double, looks at whom the article listens to and shows you where each point comes from.

    • If the article is a source for a case in the registry, we check that it names the person with their name and gender. The legal name is flagged without ever showing it.
    • More rules against revictimization (alcohol, “jealousy”, criminal record, “met him on an app”) and against morbid details of the body.
    • What is in the headline counts double: it is what most people read and what gets shared.
    • We look at whom the article gives voice to: if only the authorities speak and nobody else (family, collectives, specialists), we tell you.
    • Each result has the breakdown: how much each thing deducts or adds, from 100 to the final rating.
    • Next to each flag there is an “Is this flag wrong?” button: the team reviews each report and adjusts Turing.
    • We read the full text of long articles and PDFs (before, only the beginning).
    • When the same text appears in several outlets, it counts only once in the Observatory’s overall figures.

  14. Sensationalism or context, and the year of each text

    The Analyzer distinguishes a sensationalist text from one that gives context (data, background or analysis of how the press reports), and takes into account the year it was written.

    • Each result says what type of text it is: news article, context piece, reference document or sensationalist text.
    • If a context piece describes how the press labeled a crime (“crime of passion”, for example), that expression no longer deducts.
    • The year matters: in texts from 2014 or earlier, the terms of their time are explained and we tell you how they are said today, but they do not deduct.
    • If you paste a text or upload a PDF without a date, we ask you what year it is from before rating it.
    • Each result and each PDF state which Turing version analyzed them.
    • In possible cases, “Distrito Federal” and “D.F.” are recognized as Mexico City.

  15. Events in Mexico and foreign press

    When an article, report or study talks about LGBTTTIQANB+ people who were murdered, disappeared or attacked, the Analyzer places each of them in Mexico or abroad.

    • In reports, studies and legal documents, the terms of their time (like “a transsexual” in a text from 2001) are flagged with an explanation and how it is said today, but no longer deduct points.
    • Each flagged expression explains why we marked it.
    • If the events took place outside Mexico, the article is saved as foreign press: it counts for statistics, but not as a Memoria y Resistencia case.
    • From reports, studies and legal documents we take each person with their date, place, age and an excerpt of the text, so the team can register them without copying by hand.
    • The page of each article in the Observatory indicates whether the events took place outside Mexico.

  16. Turing 1.2

    Turing understands context

    The Analyzer now distinguishes whether an offensive expression is used by the writer or whether the text quotes it, denounces it or puts it in other people’s mouths.

    • Direct quotes, attributed words (“they called him…”, “according to the police…”) and expressions the text names to point them out as an offense no longer deduct points.
    • It recognizes human rights reports (Amnesty International, CNDH, IACHR, UN and others) and legal documents, in addition to academic studies.
    • In the underlined text, each expression explains whether it is a quote, a term or other people’s words.
    • We re-rated all the Observatory's articles with the new rules.
    • You can create an account (optional) so we can notify you of each improvement.

  17. Turing 1.1

    Recognizes academic studies

    Articles from journals and repositories (Redalyc, SciELO, theses) are no longer rated as press articles: what they quote or analyze does not count in the rating.

  18. “Did you see an article?” in the Observatory

    From any page of the Media Observatory you can paste a link and the Analyzer opens already working. The list of outlets and accounts is now paginated.

  19. Analyzes PDF documents

    The Analyzer reads links that lead to a PDF and also PDFs you upload from your device (up to 20 MB). New “Paste” button to bring in the link with one tap.

  20. Underlined text and PDF report

    Each analysis shows the text with the expressions underlined and a bubble explaining why. You can download the analysis as a PDF with a QR code.

  21. The Article analyzer and the Media Observatory are born

    Paste the link to an article or post and get a rating from 0 to 100 on how it names LGBTTTIQANB+ people. Each article on the topic is added to the Media Observatory.